Content status and review history
The site owner accepted the staging documentation and articles for publication on 4 October 2026. This report preserves their verification dates and technical review history alongside the accepted publication status.
- Current
- 189
- Awaiting publication review
- 0
- Archived
- 2
- Maintained refreshes
- 141
- AI pass
- 188
- AI fail
- 0
- AI pending
- 0
Technical review history
The initial technical screening used GPT-5.6 Sol at xhigh reasoning effort. Later corrections record their own date, model, and reasoning effort; these are correction checks, not independent editorial approval. Publication approval is recorded separately from these automated checks. Results show the latest check; earlier findings remain visible.
| Page | Status | AI Review | Last verified | Primary sources |
|---|---|---|---|---|
| Historical Codec Wiki privacy policy | archived | PassReview basis· GPT-6 Astra · medium · Correction check Preserves the old policy inside explicitly historical disclosure, corrects title/description, and directs current readers to Encode Wiki privacy; does not reuse legacy hosting/analytics claims as current policy. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh Despite archived metadata, the body still describes the current site in the present tense as GitHub Pages using Plausible in Singapore; Encode Wiki is instead served from its current infrastructure with self-hosted Umami, so the policy is misleading rather than a clearly historical record. | 9/7/2026 | 2 |
| Historical Codec Wiki terms of use | archived | PassReview basis· GPT-6 Astra · medium · Correction check Clearly limits the preserved terms, comments, cookies and liability text to historical Codec Wiki operation and links current content licensing rather than presenting old terms as binding here. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The body presents binding, present-tense terms for the old Codec Wiki and discusses nonexistent on-site comments while not clearly limiting itself to historical operation; it should not serve as current terms even though the CC BY-SA summary is broadly accurate. | 9/7/2026 | 2 |
| 5fish/SVT-AV1 | current | PassReview basis· claude-opus-5-5 · high · Correction check Snapshot moved from v2.3.260819 to the current v2.3.261005 (1170a8d5). Added --psy-bias-dg, --balancing-mg-dist-q-bias, --lambda-scale-factors and --tier. Replaced the 255/127 'automatic sentinel' defaults with the values the encoder resolves and the psy-bias levels that change them; --lineart-/--texture-psy-bias are off when omitted, and setting one sets the other to min(value, 2) (verified in a test encode). Fixed --balancing-noise-level-q-bias (1; the value is a log2 exponent, help's −2 is wrong), --balancing-mg-dist-q-bias (3, which also forces --psy-bias-dg/--enable-dg on), --lineart/--texture-variance-thr (5.5, not the internal 44), --sharpness 2, --min-qp 1, --ss −1/0/1, and noted that --enable-hdr is parsed but unused. Regrouped (psy-bias, balancing, DLF/CDEF bias groups) and rewrote all 194 descriptions. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The v2.3.260819 tag, four primary controls, 2D-content positioning, fork-specific command warning, and evaluation boundaries match the project documentation. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks the complete fork-specific registered inventory (190 rows at 2.3.260819 revision 59b8033e6e3a), default initialization and authored parameter explanations against pinned source. Adds practical native/pass-through combinations without borrowing mainline defaults; fork examples remain source-checked rather than newly binary-tested. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Page was a generic evaluation checklist that said to 'use --help for ranges/defaults', had no real starting values, and was pinned to v2.3.260819 (no longer a published release). Rewritten around upstream's own guidance at v2.3.261005: maintainers (Akatsumekusa, 5fish), SVT-AV1-PSY 2.3.0 lineage (why versions are 2.3.x and mainline 3.x/4.x options like --avif/--lossless/--rtc are absent), downloads and asset order, upstream's example command, preset advice (2 best; avoid 0), CRF bands, psy-bias starting points, what happens when psy-biases are omitted (both 0; one set → other = min(value, 2)), PSY-style tune table (3 = Subjective SSIM, 4 = Still Picture, default 0), verified default differences, and a fork-option table. Removed the arbitrary '--lineart-psy-bias 3 --texture-psy-bias 0' isolation experiments. Bound tables kept verbatim. Independent review: Added the practical consequence of 5fish's scd default: unlike mainline it inserts keyframes at scene cuts (pd_process.c sets idr/cra on scene_change_flag; --min-keyint limits spacing). Verified: v2.3.261005 is the latest non-prerelease (2026-10-05); psy-bias omission rule (both 0, or the other = min(value, 2)) in enc_handle.c; qm-min 8/9 and sharpness 2 resolution; preset 4 and tune 0 defaults; README preset advice (2 best, avoid 0), example command and build line. Coordinator: removed inline code formatting from a table header row, which failed WCAG colour contrast on the themed table headers. | 10/8/2026 | 6 |
| 7z | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote with 7z commands (dictionary, solid, threads, AES with header encryption), install note (p7zip unmaintained, 7zz), 7-Zip's read-only .zst support from 24.01 and pitfalls (solid extraction, Unix metadata). Independent review: Example used -ms=on -mmt=on, both defaults (ritual); now '7z a -mx=9 -md=256m'. Table row changed to -mmt=N (all cores by default). Commands tested with 7-Zip 26.03 (latest 26.04 per GitHub release): a/x/l/t, -ms=100m, -p -mhe=on. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is an unsourced help-wanted draft whose fixed compression-percentage claim, platform-support claims, and performance generalizations are not scoped to data, methods, versions, or tools. Earlier pass: 2026-09-07GPT-6 Astra · medium Distinguishes format and utility, removes fixed savings/platform guarantees, and adds tested 7-Zip creation, testing and restore commands with solid/encryption tradeoffs. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked 7z format capabilities, reference implementations, practical options and resource/integrity boundaries against primary sources. Binary unavailable; command shapes are source-checked, with no executed round-trip claim. | 10/8/2026 | 2 |
| AAC | current | PassReview basis· claude-opus-5-5 · high · Correction check Native AAC per aacenc.c at n9.0.2: only -aac_ms is tri-state (auto); -aac_is/-aac_pns/-aac_tns/-aac_pce are plain booleans (1/1/1/0); coders are twoloop (default) and fast only — anmr was removed although encoders.texi still documents it. libfdk_aac: real default-bitrate rule, -signaling/-latm/-header_period behaviour, valid -frame_length values. fdkaac 1.0.9 rows rewritten; -q:a row added to the generic table. Coordinator: fdkaac VBR 4 is about 148 kbit/s for stereo per FDK's aacenc_lib.h table, not 64 kbit/s per channel. Coordinator: fixed the fdkaac VBR comment, and table header code now renders with readable contrast after a global CSS fix for code in header cells. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page correctly distinguishes AAC profiles, encoder implementations, transports, and deployment-specific licensing, and its commands and cautions match current FFmpeg and FDK documentation. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Adds all native AAC, FFmpeg FDK and fdkaac parameter defaults, aliases and interactions. Native AAC/standalone FDK examples passed strict decoding; absent FFmpeg FDK wrapper and proprietary Apple runtime are explicitly source-only. Separates LC/HE/xHE profiles and dependent bitrate defaults. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds Audio Object Type identifiers, lossy PCM precision boundaries and profile-specific playback guidance. Checks xHE-AAC platform integration separately from HTML, WebCodecs and WebRTC support; retains the prior native/FDK/fdkaac parameter inventories without claiming new profile execution. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Page was a ~4,400-word parameter dump with licensing/precision disclaimers, 'Checked' notes and an execution-check narrative, and almost no guidance on which encoder to use. Rewritten as a practical guide (profiles and when to use each, encoder ranking, FDK non-free caveat, tested FFmpeg/fdkaac starting commands, containers, gapless) with the bound parameter tables moved under 'Full option reference'; restored the xHE-AAC anchor and the community encoder guidance from the original import. Independent review: libfdk_aac licensing overclaimed ('any build with --enable-libfdk-aac is non-free'). FFmpeg 9.0.2 configure lists it in EXTERNAL_LIBRARY_NONFREE_LIST, enforced only with --enable-gpl; rewrote to: GPL builds (x264/x265) need --enable-nonfree and are unredistributable. Apple encoder profile column claimed 'HE-AAC v2 on macOS only' (unverified); qaac 3.07 has no v2 mode. Now lists LC/HE-AAC/LD/ELD and notes aac_at requests HE-AAC v2 (audiotoolboxenc.c). AAC-LD/ELD row said FDK is the only common encoder; Apple's encoder (aac_at) also maps LD/ELD. Fixed. 'FFmpeg's wrapper only supports VBR with LC' replaced with FFmpeg's own wording (libfdk_aac VBR experimental, dependable with LC). Native aac: added that -q:a VBR exists but is uncalibrated (tested: -q:a 0.5/1/2 gave ~68/112/183 kbit/s on test signal); recommend -b:a. Confirmed -profile:a aac_he fails ('Profile not supported'). Encoder ranking now attributed to Hydrogenaudio community consensus; native 'fine at 160 and up' softened to 'give it 160-192 kbit/s'. Dropped the odd 'encode the whole album track by track' gapless instruction. Bound row note from J1 checked: aac_is/aac_pns/aac_tns/aac_pce are AV_OPT_TYPE_BOOL with range -1..1, so 'auto'/-1 is accepted; the bound rows are correct. | 10/8/2026 | 9 |
| ALAC | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Removed the test narrative, the sample-hash ritual and the browser-support hedging. Now says when to use ALAC, the working FFmpeg command, that FFmpeg writes only 16/24-bit (32-bit sources become 24-bit), and compatibility. Independent review: Verified FFmpeg alacenc writes only 16/24-bit (s32 input encoded as 24-bit, tested), compression_level clipped to 0-2 default 2. 'Apple's own encoder, as used by the Music app, handles all four bit depths' narrowed to the reference encoder, which is what the source supports. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This is an empty help-wanted stub with no ALAC capabilities, container guidance, tooling, recommendations, or sources. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checks Apple reference precision, channel order and utility boundaries, then verifies FFmpeg n9.0.2 help/source and synthetic execution: 16-bit ALAC round-trips exactly while 32-bit input emits 24-bit ALAC with changed low bits. Separates Safari playback/recording from general browser support. | 10/8/2026 | 4 |
| AMF | current | PassReview basis· claude-opus-5-5 · high · Correction check Added descriptions for all 47 av1_amf options (-rc -1 picks cqp if any QP option is set, cbr if -b:v equals -maxrate, else vbr_peak; -preset is an alias of -quality; bitdepth 0 follows the input; preencode ignored with CQP; align modes) and removed the hedged intro sentence. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page conflates decode and encode support by listing VP9 among AMF encoders, and its cross-vendor quality rankings and fps ranges are unsupported and unqualified by GPU generation. Earlier pass: 2026-09-07GPT-6 Astra · medium Removed VP9 from AMF encode claims, identified H.264/HEVC/AV1 as hardware-dependent encode paths, and replaced vendor rankings/fps ranges with capability checks and a concrete h264_amf VBR example. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks all 47 FFmpeg 9.0.2 AV1 private option spellings, ranges/named values and wrapper defaults, plus codec-specific generic initialization overrides. Hardware recipes are independently source/help-checked, not GPU-executed; driver negotiation, automatic settings and compatibility remain explicit limits. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Replaced the SDR baseline and discovery boilerplate. Now covers AMD generations (VCN 2.0+ for H.264/HEVC, AV1 encode only RDNA 3/4 and Phoenix APUs), Windows-vs-Linux reality (Vulkan-based AMF needs amdgpu-pro, VAAPI preferred), CQP as the quality-targeting mode with H.264/HEVC vs AV1 QP scales, automatic rc selection logic from amfenc_hevc.c, and an options table. Marked block moved verbatim to the appendix. Independent review: -quality high_quality exists for HEVC too (value 15), not only H.264/AV1; fixed both the prose and the table. Pre-analysis requirement for QVBR/HQVBR/HQCBR is stated in AMD's H.264, HEVC and AV1 API docs alike, not just AV1; table updated. Added QVBR (-rc qvbr -qvbr_quality_level with -preanalysis 1) as the closest CRF-like mode instead of claiming AMF has none. Automatic rate-control selection verified in amfenc_hevc.c/amfenc_av1.c (any -qp_*/-min_qp_*/-max_qp_* -> CQP, -b:v == -maxrate -> CBR, else peak VBR); -max_qp_* added. QP ranges 0-51 / AV1 0-255 verified; replaced the unverified 'roughly four times' with 'several times'. Hardware rows reworded to AMD's own table (starts at VCN 2.0; AV1 encode on VCN 4.0/5.0: RDNA3, RDNA4, Phoenix APUs). Linux AMF via amdgpu-pro + Vulkan init verified in general_contents.texi. | 10/8/2026 | 5 |
| Antialiasing | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Was generic text about resizing with no useful antialiasing method. Now explains what aliasing looks like, how to tell it from combing/halos, and shows vsaa NNEDI3/EEDI3/SangNom antialias and based_aa with a CPU supersampler, with the key parameters and a checking section. Independent review: Verified AADirection values, based_aa defaults (rfactor 2.0, mask_thr 60, ArtCNN supersampler, show_mask) and `vsscale onnx download ArtCNN` syntax. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is only a Help Wanted admonition and contains no antialiasing concepts, filter choices, usage, or cautions. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the Help Wanted stub with aliasing/combing diagnosis, a bicubic downscale example, edge-directed filtering limits, and native-size motion checks. | 10/8/2026 | 2 |
| aom-av1-lavish | current | PassReview basis· claude-opus-5-5 · high · Correction check The snapshot had taken the CONFIG_REALTIME_ONLY default_extra_cfg: --cpu-used 7 (normal builds: 3), --deltaq-mode 0 (normal: 1), cost-update frequencies 3/off (normal: 0/superblock); fixed all of these. Other defaults corrected from the fork's normal-build config and usage cfg: --kf-min-dist 12 and --kf-max-dist 240 (were mainline 0/9999), --gf-max-pyr-height 4 (was 5), --enable-chroma-deltaq 1, --enable-dnl-denoising 0, and realtime undershoot/overshoot/buffer values identical to good-quality. Rows flag where the fork's --help is wrong (bit depth claimed 10, code 8; qm-min/max claimed 8/15, code 5/9; force-video-mode and quant-b-adapt claimed 1, code 0). Fork options grouped as 'Options added by aom-av1-lavish' (21) with ranges from validate_config: --luma-bias -5..5, --quant-sharpness -7..7 (sign has no effect in av1_quantize.c), chroma offsets -63..63, multipliers 1..1000, butteraugli-intensity-target 0..2000, etc. Fork-specific behaviour documented: enable-cdef 0-2, deltaq-mode 0-5 (no Variance Boost), tune/tune-content values with build requirements, --enable-chroma-deltaq only acting when both --chroma-q-offset-u/v are non-zero, lag up to 128. Intro corrected: Endless_Merging (83500ccf) adds 8 options and does not drop --force-video-mode (checked against its aomenc.c); default branch is based on libaom 3.4.0. optionCount 217, verifiedAt 2026-10-08. Prose fixes by T2: --luma-bias range/direction and --tune build requirements in "Fork-specific options"; branch base versions (Endless_Merging 3.6.0, default branch 3.4.0) in the status paragraph. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Coordinator: removed --enable-chroma-deltaq=1 from the community parameter set; it restates the fork's default and has no effect without both chroma offsets.
Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This archived-fork page contains stale and risky build instructions, unsupported universal tuning recommendations, an incorrect note naming aom-psy101, and advice to copy guessed Homebrew libraries into /usr/local. Earlier pass: 2026-09-07GPT-6 Astra · medium Removed stale universal tuning, incorrect fork naming, and guessed Homebrew library-copy instructions; pinned commit 52f6a00b, accurately described dormant rather than API-archived status, and added an isolated static build and AV1 baseline. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks the complete fork-specific registered inventory (217 rows at revision 52f6a00b4833), default initialization and authored parameter explanations against pinned source. Adds practical native/pass-through combinations without borrowing mainline defaults; fork examples remain source-checked rather than newly binary-tested. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Status was wrong ('archived' frontmatter, 'dormant' prose pinned to the default branch) and the page had untested build steps and generic 'evaluate changes' filler. It now states plainly that the fork is unmaintained: default branch last changed 2022-11-20, Endless_Merging 2023-03-04, repo last pushed 2023-11-14, based on libaom 3.6.0, not archived. Added a branch table and tested build instructions for Endless_Merging with the two patches current toolchains need (NASM 3 CMake check, GCC 14 missing include from PR #1). Restored the original community Av1an settings (lag 64 is valid in lavish, defaults removed) and tips (butteraugli-resize-factor, intensity target, tune-content=psy caveat), and added a source-verified table of fork-specific options and tunes, including which tunes need Butteraugli or VMAF builds. Linked to mainline for shared options and moved the bound table to an appendix. Independent review: Status dates verified via GitHub API: Endless_Possibility 2022-11-20, Endless_Merging 2023-03-04, repo pushed 2023-11-14, not archived; libaom 3.6.0 dated 2023-02-03. Branch table missed opmox/mainline-merge, the branch behind the 2023-11-14 date (7 commits ahead of Endless_Merging, with a fork-specific --deltaq-mode=6 unrelated to mainline Variance Boost); added a row. Rebuilt Endless_Merging in /tmp with exactly the two documented sed patches (both apply); build succeeds with GCC 16 / NASM 3 and reports 'Psy v3.6.0'. Community parameter set encodes successfully. Removed --disable-kf --kf-max-dist=9999 (Av1an adds them) and --bit-depth=10 (Av1an pipes yuv420p10le) from the Av1an parameter set; intro sentence adjusted. Fork option table checked against arg_defs.c at 83500ccf (quant-sharpness −7..7, loopfilter-sharpness, luma-bias, dq-modulate, delta-qindex-mult, butteraugli resize/intensity, tune-content psy/animation/default_no_screen); lag max 128 and cpu-used default 3 confirmed. | 10/8/2026 | 6 |
| aom-psy101 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrote the native table (211 options) with the shared aomenc descriptions plus psy101-specific wording. Defaults checked against the fork's default_extra_cfg, usage cfg and quant_common.h: cpu-used 4, tune ssim, arnr 15/2 (strength range 0-4), enable-cdef 0, enable-qm 1, qm-min 0 (4 all-intra), chroma-deltaq 1, quant-b-adapt 1, denoise-block-size 16, kf-min-dist 12 (was listed as mainline 0), lag up to 128, auto-tiling 3. Flags the fork's wrong --help for --enable-dnl-denoising (claims 0; normal builds default to 1). Documents psy101-specific behaviour: chroma delta-q offsets 0/+3/+6 by subsampling outside all-intra iq/ssimulacra2, --tune-content=psy101 overrides (restoration, tx64, 64x64 SB, pyramid 4, deltaq 2/6), --fast-decode levels, --auto-tiling minimum tile sizes, low-complexity decode without the sharpness condition, HDR delta-q without the 10-bit condition. Eight fork-only options grouped separately with ranges. verifiedAt 2026-10-08. Prose fixes by T2: --kf-min-dist and all-intra qm-min added to the defaults table; --ssim-rd-mult tune list corrected. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page explicitly remains incomplete, supplies only unpinned source-build instructions, and offers no reviewed release, lifecycle status, option reference, usage, or validation guidance. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the incomplete unpinned build-only page with psy101 commit 74890d72, source-revision lifecycle scope, isolated CMake build, a minimal aomenc command, and version-specific option and decoder checks. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks the complete fork-specific registered inventory (211 rows at revision 74890d72b185), default initialization and authored parameter explanations against pinned source. Adds practical native/pass-through combinations without borrowing mainline defaults; fork examples remain source-checked rather than newly binary-tested. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Status was wrong ('archived') and the page was the same generic build/baseline/evaluate template as lavish. The original was an install-only 'Under Maintenance' stub. Rewritten with maintainer, location and status (upstream libaom merged 2026-06-02, binary reports 3.14.1-174-g74890d72b; fork-specific changes last 2025-02-21; no releases or docs besides a 2023-12-03 tag). Added a tested build and encode command, a table of defaults that differ from mainline (cpu-used 4, tune ssim, QMs on with qm-min 0, chroma deltaq, quant-b-adapt, CDEF off, arnr 2/15, lag up to 128, auto-tiling 3), and source-verified descriptions of tune-content psy101/psy, the image_perceptual_quality tune, luma-bias, auto-tiling, fast-decode and the SSIM/VMAF RD controls. The bound table was moved to an appendix. Independent review: Status verified via GitLab API: psy101 head 74890d72 (upstream merge 2026-06-02), last fork-specific commits 2025-02-21, regular upstream merges since; not archived. Defaults table verified in the non-realtime default_extra_cfg and arg_defs at 74890d72: cpu-used 4 (0..9), tune ssim, enable_qm 1 with qm_min 0, chroma deltaq 1, quant_b_adapt 1, CDEF 0, arnr 2 (max 4)/15, MAX_LAG_BUFFERS 128, denoise block 16, auto-tiling 3. tune-content=psy101 overrides (deltaq 2 in two-pass, Variance Boost in all-intra, no loop restoration, 64x64 SB) confirmed in av1_cx_iface.c. Removed redundant --cpu-used=4 from the starting command (it is the fork's default). Built the fork in /tmp and ran the command successfully. | 10/8/2026 | 5 |
| aomenc | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrote both appendices (FFmpeg n9.0.2 libaom-av1 private options, 58; native aomenc 3.15.1, 204) with sourced one- or two-sentence descriptions, regrouped into topical <details> sections (rate control, GOP, lookahead/alt-refs, delta-q, quantisation, tiles, partitioning, intra, inter, transforms, loop filters, resize, film grain, entropy, signalling). Fixed junk defaults in av1-aomenc.json ('features' for --frame-parallel/--error-resilient, 'boost' for --frame-boost, 'pyramid'/'specified' for the fixed-QP options) and replaced filler descriptions. Source findings now in the table: --static-thresh and --frame-boost are accepted but unused by the AV1 encoder; --timebase is overwritten by 1/fps; --fixed-qp-offsets only switches on --use-fixed-qp-offsets=1 and its values are ignored; --enable-chroma-deltaq outside --tune=iq/ssimulacra2 only raises chroma qindex by 2; good-quality --cpu-used accepts up to 9 but speed features stop at 6; --noise-sensitivity needs CONFIG_AV1_TEMPORAL_DENOISING and speed >= 8; vmaf_saliency_map needs CONFIG_SALIENCY_MAP. FFmpeg rows now show the effective libaom default for -1 values (e.g. -lag-in-frames -1 (libaom: 35)), note -crf defaulting to Q mode at 32 without -b:v, -crf 0 lossless, first-pass lag 0, -threads defaulting to all cores, and what -aom-params cannot set. verifiedAt set to 2026-10-08 in both snapshots; arg_defs.c added to sourceFiles. Prose fixes by T2: chroma delta-q row in "Options that matter" and the -aom-params bullet now match the source.
Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The v3.14.1 version, interfaces, build shape, format boundaries, conservative examples, and reproducibility cautions align with the tagged libaom sources. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks libaom 3.15.1 native registered options and all 58 FFmpeg 9.0.2 private options, with versioned defaults and distinct namespaces. Native/wrapper sample workflows passed strict decode/frame-count checks; copied help explanations replaced by original summaries. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Replaced the hedged prose, 'conservative example', reproducibility checklist, licence note, testing narrative and four demo commands with a practical guide. It covers current use (Av1an chunking, AVIF, forks, comparison with SVT-AV1), a tested constant-quality command, piping, an Av1an parameter set, a quality-option table, film grain, FFmpeg libaom-av1, AVIF and the fork status list. Corrections: --lag-in-frames=64 from the original guide is rejected by libaom 3.15.1 (max 48); piping into the default two-pass mode writes an empty file, so --passes=1 or explicit --pass/--fpf is needed; FFmpeg's -crf alone already selects Q mode (no -b:v 0 needed) and the wrapper's default cpu-used is 1; tune=iq/ssimulacra2 work outside all-intra since 3.14; added Variance Boost, CMake ENABLE_APPS-era build and the order-dependence of tune overrides. The bound FFmpeg and native reference tables were moved verbatim to an appendix with short intros. Independent review: Verified against libaom v3.15.1 (still the latest tag): MAX_LAG_BUFFERS 48 (lag 64 rejected), tune=iq/ssimulacra2 enable QMs, sharpness 7, QM-PSNR, adaptive CDEF, chroma deltaq and deltaq-mode 6 (handle_tuning), and work in good-quality/realtime since 3.14 (CHANGELOG); tune=iq and Variance Boost came from SVT-AV1-PSY in 3.12. --threads row said default '0 (library default)'; in libaom g_threads=0 means single-threaded (ethread.c max_threads<=1 → 1 worker). Changed to '0 (one thread)' and added that aomenc does not thread unless --threads is set. Piping into default two-pass: confirmed it fails with 'WebM writer finalization failed' and a 110-byte file; reworded the sentence to say so instead of 'cannot rewind'. --passes=1 works. Av1an 0.5.2 merges its own aom defaults (--disable-kf --kf-max-dist=9999, threads, cpu-used, cq-level) into user -v strings (settings.rs), so --disable-kf --kf-max-dist=9999 were dropped from the Av1an command and the text now says Av1an adds them; matching table-row note updated. Confirmed FFmpeg n9.0.2 libaom-av1: -crf alone → AOM_Q, -crf with -b:v → CQ, cpu-used default 1, no settings → CRF 32. Community parameter set compared with the current codecs.wiki aomenc page (still the same lag-64/deltaq-0/qm/arnr-1 set); the page's set is that set minus defaults and the invalid lag 64. Tested: constant-quality file encode, pipe with/without --passes=1, Av1an parameter string run directly through aomenc, FFmpeg libaom-av1 command, avifenc tune=iq command. Coordinator: simplified the avifenc example to match the AVIF page (libavif 1.4 with libaom 3.13+ already defaults to tune=iq for stills) and tested it with local avifenc. | 10/8/2026 | 10 |
| Audio and subtitles | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Cut ~40%: removed -n, 0:V:0, repeated 'not a policy/not a guarantee' caveats and the testing narrative. Now covers stream-specifier numbering, -map 0 -c copy remux, per-stream Opus bitrates with the 5.1(side) fix scoped via -filter:a:N (a bare -af would upmix other tracks), language/disposition/forced flags, a subtitle delivery table, MP4 mov_text + faststart, attaching fonts with mkvmerge, burn-in for text (subtitles filter loads MKV-attached fonts, si=) and bitmap subtitles (overlay), and drift vs constant offset with -itsoffset / mkvmerge --sync. Independent review: The 5.1(side) fix was described in prose only, never shown as a command. Added a complete tested command: -map 0 -c copy -c:a libopus with per-stream bitrates, -filter:a:1 aformat=channel_layouts=5.1 and -mapping_family:a:1 1. Confirmed both claims: -mapping_family 1 alone still fails ('Invalid channel layout 5.1(side) for specified mapping family 1'), and a bare -af aformat=channel_layouts=5.1 upmixes a stereo track to 5.1. Added why mapping family 1 is worth setting (surround encoder with masking). Verified remux, selective map, metadata/disposition, MP4 mov_text + faststart and subtitles-filter burn-in commands locally. Coordinator: added overlay=shortest=1 to the bitmap-subtitle burn-in command after reviewer R6a showed the unbounded output fails to mux. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The guide correctly uses explicit and optional stream mapping, distinguishes streamcopy from conversion, warns about container and subtitle compatibility, and requires stream-level and playback validation. Earlier pass: 2026-09-08GPT-6 Astra · not recorded Adds track inventory, output-relative mapping, languages/dispositions, attachment/chapter retention, text delivery, fonts, burn-in and offset diagnosis; verifies two audio tracks, subtitles, attached font, chapter retention, mov_text and a 250 ms audio offset with FFmpeg 9.0.1. | 10/8/2026 | 7 |
| Audio codecs overview | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Overview was a capability matrix full of hedging and no recommendation. Replaced with a which-codec-to-use table, a short comparison and browser notes. Independent review: Codec chooser agrees with /guides/choose-a-codec (Opus default lossy, AAC-LC for compatibility, FLAC lossless, copy source tracks); no change needed there. MP3 row said 'obsolete in quality terms'; replaced with the accurate trade-off (needs more bitrate than AAC/Opus, plays everywhere). Browser note lumped Dolby with Speex/WavPack as needing an app decoder; AC-3/E-AC-3 play where the OS supplies a licensed decoder (Apple platforms, some Windows/TV browsers). Fixed. Safari 18.4 Ogg Opus/Vorbis claim verified against the WebKit 18.4 post (macOS 15.4, iOS/iPadOS 18.4, visionOS 2.4). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The overview accurately separates codec, carriage, loss model, and immersive-audio concerns and gives useful, appropriately qualified selection guidance. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds Speex and Dolby coverage and a sample/channel comparison for every listed audio codec. Distinguishes source precision, coded bandwidth, multistream channels, lossless/hybrid modes and spatial rendering; dates browser carriage and OS dependencies. | 10/8/2026 | 4 |
| Aurora1 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Cut ~1,000 words of generic procurement and evaluation checklists. Now states only what Visionular publishes: commercial AV1 encoder, SDKs (x86/ARM/mobile) and AuroraCloud (live/VOD/RTC, free trial), use cases, plus the vendor's claims (30–50% lower bitrate, several times faster, MSU wins) attributed and qualified. Points to open alternatives. Independent review: The vendor page names video conferencing and screen content but not screen sharing; changed the use-case wording accordingly. Other claims (founded 2018, SDKs, AuroraCloud free trial, 30–50%, 'several times' faster, MSU wins, real-time presets) match the vendor page. Earlier pass: 2026-09-08GPT-6 Astra · not recorded Replaces a one-paragraph vendor stub with sourced SDK-versus-cloud distinctions, a controlled trial/comparison plan and delivery acceptance checks. Vendor claims remain attributed; no licensed SDK, version-specific CLI, independent benchmark or purchasing recommendation is asserted. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Retains controlled SDK-versus-cloud evaluation guidance and adds the explicit lack of a complete public versioned parameter/default reference. No supplier-specific syntax or defaults are imported from other AV1 encoders, and no licensed SDK/account execution is claimed. | 10/8/2026 | 1 |
| Auto-Boost Av1an Portable | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The release field is already stale: upstream published 3.4.6 on 2026-08-23, one day after the page's reviewed 3.4.5 release. Earlier pass: 2026-09-07GPT-6 Astra · medium Updated release date, tag and source commit to 3.4.6 and retained explicit bundled-dependency and quality-evaluation limits. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Updated stale 3.4.6 to 3.4.7 (September 2026), removed the reproducibility/safe-use checklists and caution box, and explained what Auto-Boost actually does (fast pass, per-scene SSIMULACRA2/XPSNR scoring, CRF zones), crediting nekotrix's original auto-boost-algorithm and Auto-Boost-Essential, followed by the portable bundle's real usage (video-input folder, bundled .bat files, bat-builder, forks, benchmark files, resume behaviour). Independent review: Verified against the portable README and releases: 3.4.7 published 2026-09-29, GPL-3.0, video-input/video-output, bundled autoboost-*/av1an-batch-* .bat files and bat-builder, forks 5fish/essential/hdr/custom, workercount-config.txt benchmark, resume behaviour with final pass restart, Linux fork link; algorithm description and Auto-Boost-Essential pointer match nekotrix/auto-boost-algorithm README (2.0 SSIMULACRA2, 2.5 adds XPSNR). | 10/8/2026 | 3 |
| Autocompressor | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the 'reproducible target-size trial' procedure, ffprobe validation commands and data-policy hedging with a short description of the hosted service and how to use it: input limits (2 GiB, 1-hour link limit), size presets 1-500 MiB plus emote/sticker, format families and their compatibility, effort tiers, 24-hour file deletion, and links to FFmpeg and the Discord page. Independent review: Verified against autocompressor.net and /about: 2 GiB input, yt-dlp links rejected over 1 hour, presets 1-500 MiB plus emote/sticker, custom sizes paid, Ultra/Extreme effort paid, 24-hour deletion, Auto-Rez Media Technologies, FFmpeg/ImageMagick/Gifsicle/MozJPEG/ECT. Earlier pass: 2026-10-04Codex (runtime model identifier not recorded) · not recorded Checked v0.4 public form controls, anonymous checked/disabled states, upload/download limits and proprietary processing scope against live primary pages. Removed obsolete marketing/account-security claims; no hosted conversion, account purchase or API contract is claimed. Both local inspection commands passed on a synthetic WebM fixture, and optional stream selectors are shell-quoted following independent review. | 10/8/2026 | 2 |
| AV1 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Cut from ~6,300 to ~1,850 words. Removed the 'team review in progress' caution, the duplicated profile tables, the 'practical workflow' checklist and the 700-line migrated spec dump (partition/transform/OBU index tables); replaced them with a plain-language coding-tools section restored from the community original (Qu Pengfei overview) and linked to it. Corrected the level claim: the syntax spans 2.0-7.3 but only 14 levels are defined (2.0, 2.1, 3.0, 3.1, 4.0, 4.1, 5.0-6.3); release dates corrected to v1.0.0 June 2018 / Errata 1 Jan 2019. Added when-to-use guidance, film grain synthesis practice (encoder options checked against --help; metric caveat), verified hardware decode/encode generations, browser summary, licensing (AOMedia patent license, Sisvel pool, Avanci Video, closed EU inquiry) and pitfalls. Independent review: Verified Annex A: seq_level_idx spans 2.0-7.3, limits are defined only for 2.0, 2.1, 3.0, 3.1, 4.0, 4.1 and 5.0-6.3 (14 levels); High tier only from 4.0; 4.0/5.1/6.0 MaxDisplayRate matches 1080p30/2160p60/4320p30. Profiles match. Release tags: v1.0.0 25 June 2018, v1.0.0-errata1 8 Jan 2019. Hardware table had an unconfirmable encode column ('Check the device' for Apple/Android). Made it decode-only and added MediaTek Dimensity 1000 (first AV1 SoC decoder). Intel encode is now 'Arc and Core Ultra integrated graphics' (Meteor Lake media engine has AV1 encode). Removed claims that Edge decodes AV1 everywhere; only Chrome and Firefox are stated (Edge's built-in AV1 support could not be confirmed from a Microsoft source). Film grain options were checked against SVT-AV1 v4.2.0, aomenc v3.15.1 and rav1e 0.8.1 --help. Added that SVT-AV1 --film-grain does not denoise the coded picture unless --film-grain-denoise 1 (default 0), the 1-50 range, and that all three load grain tables. Restored coding-tools section checked: 56 directional angles, 10 partition types, 16 transform combinations, 64-pt transforms coding only 32x32, CDEF origin, IBC restrictions, super-resolution order all correct. WebM guideline statement verified (guidelines still list only VP8/VP9). Licensing verified: AOMedia patent license, Sisvel AV1/VP9 pool (2019), Avanci Video covers AV1/HEVC/VVC/VP9 for streaming services.
Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The current overview is strong, but the user-facing migrated technical reference contains material errors, including calling CAVLC an arithmetic coder and implying HEVC uses it; HEVC entropy coding uses CABAC, while CAVLC is an AVC variable-length coding mode. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected migrated entropy reference to distinguish AVC CAVLC from CABAC in AVC/HEVC and AV1 multi-symbol arithmetic coding. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked professional-profile depth/chroma restrictions, SVT-AV1 v4.2.0 4:2:0 input limits, rav1e precision/chroma coverage, and decoder-versus-hardware/browser distinctions. Preserved the separately labeled migrated coding-tools material for human editorial review. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Corrected inherited normative errors: distinct AV1/VP9 nominal intra angles and Paeth/TrueMotion names, CDEF 8x8 direction versus 64x64 strength regions, missing INTER_3 transform-set column, transform-equals-block cases, tile width/area and dependency meaning, segmentation delta/skip semantics, four spatial ID values, OBU_TILE_LIST type8, and NAL expansion. Preserved credit and the explicitly pending human editorial review of the migrated technical appendix. | 10/8/2026 | 12 |
| AV1 Encoding for Dummies | current | PassReview basis· GPT-6 Astra · medium · Correction check Preserves credited 2023 command and illustrations as historical material, replaces obsolete installers/source-line patches, fixes WSL2/10-bit/CQ/tiling explanations and stream handling, and routes new work to maintained guides. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is explicitly marked outdated but remains an install-and-settings guide: it pins obsolete dependencies and abandoned forks, gives brittle source patches, and presents a September 2023 aomenc command as recommended guidance rather than an archival example. | 9/7/2026 | 5 |
| AV1 for Dummies | current | PassReview basis· GPT-6 Astra · medium · Correction check Scopes original tool ratings to the 2024 authors' survey, corrects 10-bit compatibility guarantees and encoder rankings/capabilities, records PSY sunset, and replaces stale selection prose with constraint-based guidance. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The guide makes materially overbroad recommendations—especially that 10-bit always improves quality with zero compatibility penalty—and its ranked tools, encoder capabilities, preset descriptions, and preferred SVT-AV1-PSY project are stale after major upstream changes and the original fork's sunset. | 9/7/2026 | 6 |
| AV1 with SVT-AV1 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the version-check preamble, 30-second sample ritual, CRF 28/32/36 trial loop, HDR validation essay and 'retain the recipe' section with an upstream-consistent SVT-AV1 v4.2.0 workflow: preset 4 / CRF 30 / 10-bit / tune=0 starting command, preset table, CRF range and FFmpeg's 63 cap, tune semantics (default tune is 1 = PSNR), 10-bit rationale (AV1 Main includes 10-bit), film-grain levels from upstream with the correct film-grain-denoise semantics (default 0 since v2.0.0, so grain is added on top of the undenoised source), keyframe interval via -g or keyint=10s, variance boost/ac-bias/fast-decode table, SvtAv1EncApp Y4M pipe (with the still-required -strict -1) plus mkvmerge mux and HDR flags, and a brief Av1an pointer. Independent review: Starting command was inconsistent with the SVT-AV1 page (only tune=0); now the same preset 4 / CRF 30 / 10-bit / tune=0:enable-variance-boost=1:keyint=10s command, also in the film-grain and SvtAv1EncApp examples. Film-grain example used film-grain-denoise=1 while both pages describe denoise 0 (the default, upstream's higher-fidelity option) as normal; example now uses the default. Keyframe section rewritten: default is ~5 s (160 frames at 24 fps), -g has no seconds form, and added that SVT-AV1 never places scene-cut keyframes (scd ignored) with the Av1an pointer, which the guide lacked. Added FFmpeg -crf integer rounding (27.5 -> 28) and the preset 12/13 -> 11 mapping outside rtc. variance-boost-strength row now includes upstream's strength 3 case. Verified unchanged: upstream crf=30 starting point and CRF 25 VOD example, '10 x fps, max 300' hobbyist GOP rule, 2D animation film-grain 4-6 (Ffmpeg.md), grain-synthesis warning above preset 6 (enc_settings.c), -strict -1 still required for 10-bit Y4M in FFmpeg 9.0.2, SvtAv1EncApp accepts '-i -' (app_config.c). Coordinator: removed inline code formatting from a table header row, which failed WCAG colour contrast on the themed table headers.
Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The sample command and explanations match the current FFmpeg wrapper and SVT-AV1 4.0 controls; the page appropriately presents its settings as a test starting point and calls for version and decoder validation. Earlier pass: 2026-09-07GPT-6 Astra · medium Expanded practical mainline AV1 recipes and explained quality/preset selection, pipelines, stream preservation and HDR limits; 4.2 documentation is distinguished from local 4.1 smoke tests. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted four optional FFmpeg -map values so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. | 10/8/2026 | 6 |
| Av1an | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The maintained repository, v0.5.2 release, supported encoders, dependency model, target-quality caveats, and documented VapourSynth compatibility warning are accurately represented. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked v0.5.2 tagged parser defaults, SVT encoder/pixel format and mkvmerge concatenation default; distinguished moving website defaults. Local VSScript initialization fails and no successful encode is claimed. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Rewrote from a disclaimer/reproducibility page into a practical guide. Corrected errors: the page claimed the docs site lists ffmpeg as the concat default (it is mkvmerge), implied tagged GitHub releases carry Windows binaries (only the nightly 'latest' pre-release does), and gave no chunk-method fallback order. Added: verified VapourSynth problem (v0.5.2 panics with 'Failed to get VSScript API' on R73+; master works with R80), dependency table, tested SVT-AV1 (--tune 0) and aomenc commands, -a audio usage, key options table with real defaults (auto workers = min(CPU threads/encoder factor, RAM estimate), 10 s extra-split, 24-frame min scene), target-quality metrics and their plugin requirements (vmaf, ssimulacra2 and butteraugli via Vship/vszip/julek, xpsnr via FFmpeg), chunk-method table, and pitfalls (Av1an disables encoder keyframe placement so GOP length comes from extra-split; gray frames -> bestsource or lossless intermediate; RAM; resume). Restored the original community troubleshooting where still valid; dropped the obsolete mkvmerge 'ulimit -n' tip because Av1an now merges in groups of 960 chunks. Independent review: Default chunk-method order was wrong: both v0.5.2 and master pick bestsource first, then lsmash, ffms2, dgdecnv, hybrid (VapoursynthPlugins::best_available_chunk_method; --help text agrees). Page had lsmash first and bestsource fourth. Fixed in the bullet and chunk table, and kept `-m lsmash` with the reason. VapourSynth breakage threshold was R73; issue #1239 and #1227 show R73 works and R74+ (new packaging) breaks v0.5.2; fixes landed in master 2026-04-27 / 2026-05-07. Corrected. Nonexistent options: `--min-q`/`--max-q` and `--probe-slow` are not in v0.5.2 or master; replaced with `--qp-range` and `--probe-video-params copy`, and noted that probes use a fast fixed preset by default (construct_target_quality_command). Per coordinator/R3a: stated once how Av1an merges its default encoder arguments into -v (settings.rs: defaults prepended, flags you set are dropped from the defaults; skipped entirely with --no-defaults) and what remains: SVT-AV1 `--keyint 0 --scd 0 --rc 0`, aomenc `--threads=8 --cpu-used=6 --disable-kf --kf-max-dist=9999` (observed in Av1an's 'Params:' line); with --no-defaults the user must add keyframe disabling. Same code in v0.5.2. Consistent with the aomenc page. Per coordinator: confirmed XPSNR input swap. FFmpeg n9.0.2 vf_xpsnr uses input 0 as the original (perceptual weights computed from it; filters.texi: 'first input is considered a reference'), while Av1an builds `[distorted][ref]xpsnr` in metrics/xpsnr.rs (v0.5.2 and master 6487fa5); this path is used at the default --probing-rate 1. The vszip path passes named reference/distorted args correctly; VMAF order is correct. Local check on a synthetic clip: Y 36.14 dB correct order vs 33.38 dB swapped. Added one pitfall line. No upstream issue or PR exists (searched rust-av/Av1an for xpsnr). aomenc section: automatic worker count assumes 4 threads per aomenc worker regardless of --threads (determine_workers), so with --threads=2 users should set -w; added. xpsnr target row now says it scores the weakest plane (XPSNRSubMetric::Minimum). Verified unchanged: mkvmerge concat default and x265 requiring mkvmerge; --extra-split-sec default 10.0 (with -x frames taking priority); --min-scene-len 24; --pix-format yuv420p10le; -c:a copy default; aom/vpx two-pass default; photon-noise 0-64 for aom/rav1e/svt-av1; worker CPU divisors (svt 6, aom 4) and RAM estimate from total memory; target-metric names and plugin requirements; v0.5.2 tag has no assets and 'latest' pre-release carries av1an.exe (updated 2026-09-14). Ran master 6487fa5 locally: SVT-AV1 and aomenc commands, --no-defaults. Coordinator: removed inline code formatting from a table header row, which failed WCAG colour contrast on the themed table headers.
| 10/8/2026 | 13 |
| av1an-command-gen | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Identified the archived project and its historical Zig 0.11 build, with a pinned revision. Explained that generated settings are candidates requiring comparison with installed encoder controls, and required inspection of paths and quoting without eval. Preserved the original positional arguments and examples. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Cut to a short archived-tool page: what it is (Zig CLI that prints an Av1an command), archive status (last commit April 2024), argument syntax, build requirement (Zig 0.11.0) and a pointer to the current Av1an page; removed the body H1, pinned-commit build steps and 'review the generated command' padding. Independent review: Verified against the repository: archived=true, last push 2024-04-26, argument order and values, Zig 0.11.0 build to zig-out/bin, BSD-3-Clause. | 10/8/2026 | 1 |
| AV2 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Restated status plainly from AOMedia sources: v1.0.0 spec dated 28 May 2026 (announced June 2026), coding tools frozen per AOMedia, only AVM v1.0.0 reference software plus VideoLAN's early dav2d (0.0.x tags), no FFmpeg/browser/hardware support, no ISOBMFF/Matroska binding. Profile table re-checked against Annex A (all v1.0.0 profiles max 10-bit; levels 2.0-8.3) and the interoperability-point explanation corrected from Table A.3. Added a short verified list of new tools (256x256 superblocks, optical-flow refinement, TCQ, CCSO, multi-layer/multi-stream) and the sourced ~30% claim with its scope. Removed the duplicated checklist and 'final specification versus deployable ecosystem' section. Independent review: Verified on av2.aomedia.org: v1.0.0 dated 28 May 2026, current; v13 working draft (5 Jan 2026) superseded. Press release is dated June 2026. Verified 'coding tools are complete... core coding tools will not change' and 'future versions... additional profiles and metadata' in AOMedia's implementers blog post. Verified Annex A Table A.1 (five Main_*_10_IP* profiles plus Configurable, bit_depth_idc 0 or 1), Table A.3 interoperability points, and Table A.7 levels 2.0-8.3; added that High tier starts at 4.0 (spec note). ~30% figure re-attributed from a personal blog (Norkin) to AOMedia's own post, which states VMAF and PSNR with AVM; Norkin source dropped. Added AOMedia's announced 12-bit professional profile project from the press release. dav2d tags confirmed as 0.0.0 (Oct 2025) and 0.0.1 (Apr 2026); AVM v1.0.0 tag exists. Tool list (256x256 SB, optical-flow refinement, TCQ, CCSO, CCTX, IBC/palette, multi-layer/multi-stream) found in the v1.0.0 spec text. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The profile, level/tier, reference-software, and early-ecosystem guidance matches the final AV2 v1.0.0 specification and appropriately treats unverified container/HDR bindings as unresolved. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked the finalized May 2026 v1.0.0 specification, Annex A chroma/depth/IP constraints, and the official AVM GitHub release location. Distinguished a finalized specification from unverified broad browser, hardware, container and named-HDR deployment. | 10/8/2026 | 6 |
| AVC / H.264 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Rewrote around practical use: when to use H.264, one merged profile table (bit depth/chroma/where it plays; replaced two overlapping tables), a common-levels table with values checked against the level limits, plain-language coding tools, encoders (x264, OpenH264's Constrained-Baseline-only scope, JM, hardware), containers (avc1 vs avc3, Annex B, RTP, FLV), licensing (Via LA pool, Cisco OpenH264 royalty coverage) and pitfalls. Removed the validation checklist, stacked caveats and dated browser narrative. Added one tested FFmpeg command, explaining why -pix_fmt yuv420p matters (RGB input produces High 4:4:4 Predictive) and documenting that FFmpeg's libx264 -level caps refs while the x264 CLI only warns. Independent review: Wrong: page said the x264 CLI only warns on --level while FFmpeg caps refs. x264.c (lines ~1789-1792) also lowers i_frame_reference to fit the level DPB, and both produced ref=4 at 1080p level 4.1 with -preset veryslow. Both only warn on bitrate/frame-size limits (x264 set.c x264_validate_levels). Rewrote the paragraph. Tested the delivery command (High, yuv420p) and confirmed rgb24 input without -pix_fmt yields High 4:4:4 Predictive yuv444p. Level table values checked against the H.264 level limits (3.1 720p30, 4.0/4.1 20/50 Mbit/s Main, 4.2 1080p60, 5.1 2160p30, 5.2 2160p60); levels run 1-6.2. Current ITU-T edition 06/2026 confirmed. OpenH264 Constrained Baseline scope confirmed in README. NVIDIA RTX 50 H.264 4:2:2 decode/encode confirmed in matrix.
Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page explicitly calls H.264 version 15 from August 2024 the in-force edition, but ITU-T published in-force version 16 in June 2026 and marks version 15 superseded. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected the in-force H.264 edition to version 16, June 2026. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked the H.264 edition and fidelity-profile precision/chroma scope against the Recommendation, x264 build-dependent input/profiles, OpenH264 documented realtime scope, and AVC WebCodecs configuration boundaries. Browser/device support remains scoped to the tested rendition. | 10/8/2026 | 10 |
| AVI | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Condensed to what matters: no per-frame timestamps, VFR and B-frame problems, 2 GB limit/OpenDML, and how to move old AVI content to MKV/MP4; removed the long timing-fields discussion and testing narrative. Independent review: Found a real, missing pitfall: stream-copying an MPEG-4 ASP AVI with B-frames to MKV fails in FFmpeg 9.0.2 ('Can't write packet with unknown timestamp') and to MP4 writes 'pts has no value' warnings. -fflags +genpts fixes both; mkvmerge handles it directly. Added, plus the mpeg4_unpack_bframes bitstream filter for packed-bitstream DivX/Xvid. Added MKVToolNix to appliesTo and the bitstream-filter doc to sources. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The concise page accurately describes AVI's RIFF basis, legacy role, modern timing/metadata limitations, and preservation-first migration guidance. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Expanded the short overview with RIFF stream/index structure, OpenDML size boundaries, rate/scale timing, codec-versus-container precision, selected-track inspection/remux/decode and browser delivery limits. Checked final text against Microsoft and FFmpeg references; scoped successful synthetic checks explicitly. | 10/8/2026 | 3 |
| Aviator | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Replaced the obsolete SVT-AV1-PSY workflow with the reviewed SVT-AV1-HDR backend and source controls. Confirmed CRF 32 and preset 6 defaults, distinguished the old 0.6.0 tag from the reviewed revision, and separated Linux Flatpak support from untested WSL alternatives. Preserved screenshots; GUI execution was unavailable. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Fixed the version picture (Flathub ships 0.7.0 while the last GitHub tag is 0.6.0), removed WSL speculation, record-keeping advice and commit-pinned narration, and rewrote the settings as a concise list of the actual defaults from the README (preset 6, CRF 32, grain synth/denoise, Opus 80 kbit/s, Copy Audio and WebM restrictions). Independent review: Credit fixed: Aviator is by Gianni Rosato and Nate Sales (originally natesales/aviator; app id net.natesales.Aviator; natesales is the second-largest contributor). Verified SVT-AV1-HDR, preset 6/CRF 32, Opus 80 kbit/s, grain synth/denoise, volume/normalize, WebM restrictions from README; Flathub 0.7.0 (2025-08-31) vs last GitHub tag 0.6.0. | 10/8/2026 | 3 |
| AVIF | current | PassReview basis· claude-opus-5-5 · high · Correction check avifenc 1.4.2 and libavif revision verified current (help and avifenc.c hashes match). Rewrote all avifenc descriptions; fixed --qalpha default to 'same as qcolor'; libheif aom/rav1e/svt plugin snapshots moved to 1.23.6 with rewritten rows (aom tune=iq can't combine with lossless, rav1e quality/lossless are hidden setters, threads default no longer says '16 here'); heif-enc anchor link updated to the 1.23.6 heading. Independent review: rav1e quality said 'Accepted but not listed by --params'; heif_encoder_set_parameter only accepts listed parameters, so heif-enc -p quality=30 exits 5 ('Unsupported encoder parameter'). Quality is reachable only via -q. Fixed (type 'integer 0–100, -q only'). rav1e lossless said 'Accepted for compatibility'; -p lossless= is rejected, and -L warns and falls back to lossy (verified). Fixed. SVT chroma said 'Accepted but not listed by --params'; the chroma parameter and setter are compiled out (#if 0), -p chroma=420 exits 5, and the plugin always requests YCbCr 4:2:0. Fixed. SVT lossless omitted that heif-enc -L on RGB input fails with exit 5: heif-enc also sets chroma=444, which this plugin lacks (verified with a PNG). Added. Snapshot validation.completeness for rav1e and SVT updated to record these runtime checks. Checked and correct: every avifenc row sampled (qalpha follows qcolor until set, jobs=all, speed 6, progressive needs qcolor >= 10, target-size conflicts, irot/imir, repetition-count, autotiling, timescale/duration) and all aom rows (threads = hardware_concurrency capped at 64, 4 if undetectable; speed 0–9 on libaom 3+; tune=iq with lossless error). Coordinator: the appendix now links to the consolidated 'Reading option tables' page. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately distinguishes AVIF v1.2.0 from libavif v1.4.2 and gives current, careful guidance on backends, quality, lossless validation, color/HDR, animation, and layered images. Earlier pass: 2026-09-07GPT-6 Astra · medium Expanded practical guide with version-scoped recipes, controls, tradeoffs, compatibility, and validation. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Reviews the complete libavif 1.4.2 fixed frontend and libheif 1.23.5 AOM/rav1e/SVT plugin contracts against parser/default audits. Dynamic native passthrough is explicitly excluded; recipes distinguish lossless samples, precision reduction, sequence timing, progressive/grid carriage and separately tested XMP extraction. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checks AVIF v1.2.0 Baseline/Advanced brands separately from AV1 Main/High/Professional sample and chroma capabilities. Adds muxer/backend/decoder distinctions and dated still/animation browser milestones while preserving the complete prior frontend/plugin parameter references. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): The page was ~5,500 words of hedged capability tables, a 'Tested backend recipes' narrative and a stale 'Review status' banner, and it dropped the original's practical guidance (the original also recommended obsolete `-a tune=ssim -a cq-level` quality control). Rewritten as a practical guide: strengths/weaknesses, current browser/Android support, a minimal `avifenc` command with an options table, 4:4:4 vs 4:2:0, `tune=iq` (libaom >= 3.12, avifenc default for still non-RGB images with libaom >= 3.13), SVT-AV1 being 4:2:0-only in libavif's adapter, HDR/alpha/animation and pitfalls. The bound avifenc and libheif plugin tables are kept verbatim at the end under 'Full option reference'. Independent review: Removed the single-photo claims: '-q 80 is already in the mid-80s on SSIMULACRA2' and '10-bit is a few percent more efficient than 8 even for 8-bit sources'. The -d row now says the default follows 8-bit input, use 10 for HDR/high-bit-depth, and that -d 10 is common for 8-bit sources to reduce gradient banding. Starting command dropped the padding `-s 6` (default) and `-d 10`; now `avifenc -q 65 -s 4` with one line on choosing quality at a fast speed and finishing at -s 4. tune=iq description corrected from the libaom 3.12 release article and libavif CHANGELOG: default since libavif 1.4.0 for still non-RGB images with libaom >= 3.13, new quality-to-QP mapping, uses Variance Boost/adaptive CDEF/quantization matrices; alpha uses tune=psnr. Removed the unsupported 'evens out quality / screen content' description. Chroma section claimed 4:4:4 and 4:2:0 come out 'almost the same' for photos (one test); replaced with the avifenc default (4:4:4 for PNG), when to keep 4:4:4, and when -y 420 + --sharpyuv makes sense (smallest files, Baseline/Main profile hardware decoders). Restores the original wiki's 4:4:4-first community advice without numbers. SVT backend: confirmed locally (avifenc 1.4.2, SVT 4.2.0) that PNG without -y 420 fails with 'Only support 420 now', -d 12 fails, and libaom -a keys such as tune=iq are rejected; table updated. rav1e row given a reason. Pitfall 'strip or preserve ... deliberately' replaced with the tested behaviour: avifenc copies Exif/XMP/ICC and converts Exif orientation to irot/imir. Other encoders: dropped the odd 'SVT-AV1-PSY ... are the video encoders behind these paths'; links aomenc/SVT-AV1/rav1e; kept tinyavif. Verified support table against MDN BCD (Chrome 85, Edge 121, Firefox 93/animation 113, Safari 16.4 / iOS 16.1, 16.1 Ventura-only note) and Android formats page (AVIF decode Android 12+, 10-bit optional). Tested Y4M pipe animation command. Coordinator: linked the option appendix to the image encoder parameter reference. | 10/8/2026 | 9 |
| AVM reference software | current | PassReview basis· claude-opus-5-5 · high · Correction check Rechecked all 250 avmenc option spellings against v1.0.0 (966a7d7): apps/avmenc.c, av2/arg_defs.c and av2/av2_cx_iface.c, plus --help and test runs of a scratch build. The option set was already complete, but many defaults were wrong or garbage strings. Fixed: --passes is 1; undershoot/overshoot/buffer values no longer list realtime values (AVM only has usage 0, and --usage=1 fails); --psnr's bare flag means 2, falling back to 1; --timebase has no effect (avmenc always sets 1/fps); --fpf is listed in --help but never parsed, so it fails as unrecognized; --enable-tcq is 1 (every frame), not the 2 that --help claims, and is turned off from --cpu-used 2; --sframe-type is 0; --frame-hash is off; --target-seq-level-idx is 31; --vmaf-model-path no longer has stray quotes. Placeholder defaults ('metadata', 'filter', 'features', 'boost', 'frames', 'specified', 'frame.', 'format', 'S-Frame') are replaced with real values. Documented how the container is chosen: WebM is the default; an -o name ending in lowercase .ivf/.obu selects IVF/OBU unless --webm came first; --ivf/--obu and -o are applied left to right; without WebM I/O the output is OBU (all tested). --qp rejects negative values (it is parsed as unsigned and offset internally), while --min-qp/--max-qp accept down to -48 at 10-bit. 12-bit is rejected in normal builds. --noise-sensitivity, --static-thresh and --frame-boost are accepted but never reach the encoder. Hidden options (--tile-width/height, --fixed-qp-offsets, --use-fixed-qp-offsets, --cq-level, --frame-hash, --use-per-plane-frame-hash) are marked. Replaced 'Coding tools 1-5' with 14 real groups. Rewrote every description, removing help-text filler and boilerplate. Rewrote the intro and dropped the nonexistent Butteraugli build option. Snapshot: defaults and sections updated, verifiedAt 2026-10-08, sourceFiles 3, optionCount 250. Prose fix by T2: --qp is 0–255 at every bit depth; only --min-qp/--max-qp reach −48 at 10-bit. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page directs users to the deprecated GitLab repository and repeatedly names the current AV2 applications aomenc/aomdec; the canonical GitHub project uses avmenc/avmdec. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected canonical repository to AOMedia GitHub and all application names to avmenc/avmdec at v1.0.0. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Adds the 250-row source-pinned AVM 1.0.0 parameter/default inventory and example workflows while identifying this release as AV2, not AV1. Source-only verification is explicit; no AV2 binary execution, player compatibility or quality result is asserted. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Cut the version-discipline, record-keeping, container, colour and licensing boilerplate and the 'not built or run' narrative. Now gives AV2 v1.0.0 status (spec 28 May 2026, AVM v1.0.0 tag 27 May 2026, GitLab repo deprecated), a tested build and a basic encode/decode. Corrected the example: the previous commands relied on --qp without noting the default end-usage is VBR, so --end-usage=q is now required. Added pitfalls (cpu-used defaults to 0, single pass only, output format chosen by file extension). Dropped an unverified '~30% better than AV1' figure (not in AOMedia's press release). 250-row option table kept verbatim. Independent review: Verified v1.0.0 is still the only release tag (annotated 2026-05-27, 'First released version of AV2'), spec PDF dated 2026-05-28, press release mentions a planned ISO BMFF binding and dav2d; main branch still active. --qp line mentioned a −96 minimum at 12-bit, which reads as if 12-bit AV2 were usable although v1.0.0 profiles stop at 10-bit (per the AV2 codec page); reduced to the 8-bit/10-bit ranges. Linked dav2d (VideoLAN preview released May 2026) and described it as early development. Tested with the v1.0.0 build: --end-usage=q --qp encode to IVF, avmdec to Y4M, 10-bit FFmpeg pipe, and extension-based WebM output for a .mkv name; speed claim consistent (≈1.6 fps at 160×96, cpu-used 6). | 10/8/2026 | 6 |
| AVS3 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced a hedged, catalog-listing page with what AVS3 is, the verified DVB adoption (July 2022, A001r20), an implementations table and a working build/encode/decode recipe. Corrected the original's claim that FFmpeg decoding arrived in version 6 (it is in the 4.4 changelog) and cut the unverifiable hardware-decoder and SVT-AVS3 claims. Documented real build problems: the decoder/encoder binaries are uavs3dec/uavs3enc, and CMake 4 and GCC 14 need extra flags. Independent review: Wrong release facts: uavs3e has tags/releases v1.1 and v1.3 (Oct 2022), not 'no tagged releases', and uavs3d published release 1.2 on 2026-08-10. appliesTo is updated to v1.3 / 1.2. Added the important profile gap: DVB adopted the AVS3 High profile (finalised August 2021, T/AI 109.2-2021), while uavs3e/uavs3d implement only the baseline profile. Added the development history (started Dec 2017, main profile March 2019) from DVB-sourced coverage. Removed the duplicate encoder build recipe in favour of a link to /tools/software-encoders/uavs3e/, which carries the same flags and the 10-bit crash caveat. Kept a tested uavs3d build line (now path-explicit) and the encode/decode pair. Rebuilt both tools here. The encoder needs -Wno-error=incompatible-pointer-types plus the unistd include under GCC 16/CMake 4.4, and the binaries are uavs3enc/uavs3dec. The encode/decode commands ran. Dropped the unreachable AVS catalog source, pinned the FFmpeg Changelog to n9.0.2 (libuavs3d decoder and AVS3 demuxer in 4.4) and removed an 'In practice' tic. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page freezes ecosystem and hardware observations at 2024, relies on personal phrasing such as 'to my knowledge,' and omits later AVS3 parts and a sourced current implementation/capability review. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked AVS standard-family scope and uavs3 baseline implementation, 8/10-bit build behavior and optimized versus unoptimized paths. Limited chroma/deployment claims to the reviewed baseline software and explicitly separated broadcast, HDR, hardware and browser evidence. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added the upstream uavs3d CLI decode example, explained frame-level threads/logging, made default 8-bit versus COMPILE_10BIT decoder build explicit, and retained raw-layout sidecar requirements. Source-checked only, with no claimed local AVS3 conformance run. | 10/8/2026 | 5 |
| Basics of Filtering | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Was ~1800 words of hedged caveats and recorded-version boilerplate, with stale API (initialize_clip default described from an old helper, vsscale.SSIM and vsaa.Nnedi3 no longer exist). Rewrote as a short orientation: when to filter, filter order (kept the flowchart), starting script using vssource.BestSource (which already runs initialize_clip at 32-bit by default) and finalize_clip, cropdetect/Crop, resizing with current vskernels/vsaa names, trimming. Independent review: Removed the claims about a previewer crop tool: VSView (the current previewer) documents none and vs-preview is archived. IVTC note fixed: only VFM is 8-bit-only; VDecimate accepts 10/16-bit (tested). Verified finalize_clip(bits=10, clamp_tv_range=False), BestSource default bits=32, vsscale onnx download CLI. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Updated deprecated standalone vs-tools guidance to the reviewed vs-jetpack revision. Corrected 32-bit initialize_clip and 10-bit, unclamped finalize_clip defaults, repaired Python comments and imports, separated resizing alternatives and removed universal filter rankings. All nine Python examples passed syntax checks; plugin loading and filtering were not executed. | 10/8/2026 | 4 |
| Better late than never: SVT-AV1 v2.2.x Deep Dive | current | PassReview basis· gpt-6.1-sol · ultra · Correction check Identified the results as the author’s historical SVT-AV1 2.1.2 and 2.2.1 experiment. Explained the limits of its harmonic pooling when SSIMULACRA2 scores are zero or negative, without inventing how the original data was transformed. Preserved all charts, media and attribution; no experiment was rerun. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The version-scoped v2.1.x/v2.2.x comparison documents exact builds, settings, hardware, CRF coverage, and seven named samples spanning live action, games, animation, and HDR; it supplies size-aware visual comparisons, uses two independent objective metrics, and expressly discloses XPSNR, low-score, speed, and aggregation limitations. | 10/4/2026 | 4 |
| brotli | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote with CLI usage (default quality 11 and keeping input), window options, tar -I brotli, web delivery and large-window caveat. Independent review: Wrong default: the brotli CLI window is 24 (16 MiB), reduced for small files (DEFAULT_LGWIN 24 in c/tools/brotli.c v1.2.0), not 22/4 MiB. Fixed. Removed redundant -k (keep is the default). Server list wrongly implied Caddy compresses brotli; Caddy only serves precompressed .br via file_server precompressed (checked docs). Confirmed GNU tar 1.35 -caf .tar.br writes an uncompressed tar and tar -tf cannot read brotli; -I brotli works. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This short draft has no sources and uses a dated global browser-support percentage and unsupported compatibility judgments instead of documenting the Brotli format, parameters, and current HTTP content-coding behavior. Earlier pass: 2026-09-07GPT-6 Astra · medium Adds practical tested CLI round-trip, separates Brotli streams from archives/WOFF2, and replaces browser percentages with HTTP negotiation and integrity limits. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked RFC versus large-window extension, CLI quality and auto window behavior; corrected the library-default confusion using v1.2.0 source and independent file/stdin tests. Exact byte restoration passed. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Removed the final paragraph's contradictory implied content checksum; raw Brotli has none. Require source/restored comparison or an independent trusted hash. Brotli 1.2.0 encode/test/decode/cmp passed on 4 KiB bytes. | 10/8/2026 | 3 |
| Butteraugli | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced hedged text and a validation-heavy script with: what Butteraugli is (libjxl-maintained, Google repo archived), its high-fidelity design range from the authors, libjxl's definition of distance as target max Butteraugli distance (1.0 visually lossless), a norms table (max vs 3-norm vs q-norm with libjxl and Vship names), the intensity-target difference between tools (libjxl CLI 80 nits for SDR vs Vship 203 nits), libjxl butteraugli_main usage, and FFVship/Vship video scoring with a simplified script and distmap preview. Corrected the original's broken script and deprecated _2Norm usage. Independent review: Added a data-backed limitation: Butteraugli 3-norm is among the best on near-threshold KonFiG-IQA (SRCC 0.92) but no better than PSNR on heavily distorted TID2013/KADID-10k (0.66/0.54 vs PSNR-Y 0.64/0.63). Verified intensity targets per tool: libjxl butteraugli_main uses 80 nits for SDR (PQ/HLG use the image's intensity target), Vship defaults to 203; default --pnorm 3; output order max distance then 3-norm; Vship property names and deprecated _BUTTERAUGLI_2Norm alias; Google README wording (barely noticeable differences, JPEG q90-95). Tested: FFVship -m Butteraugli --json (JSON is [q-norm, 3-norm, max] per frame; stats printed for 2-norm, 3-norm, INF-norm) and the page's VapourSynth script. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The Vship example is not executable because all three print calls have unterminated strings, and it reads a nonexistent current _BUTTERAUGLI_2Norm property instead of Vship's documented 3Norm, INFNorm, and configurable QNorm outputs; the page is also explicitly incomplete. Earlier pass: 2026-09-07GPT-6 Astra · medium Repairs Python syntax and current QNorm/3Norm/INFNorm access with a single-pass example; documents intensity, color/alignment and temporal-pooling limits. Syntax and upstream API checked; GPU execution not tested. | 10/8/2026 | 7 |
| bzip2 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Short reference: block sizes, memory figures from the manual, tar -j, parallel implementations and recovery; clearly states it is mostly a legacy format. Independent review: Removed padding line 'bzip2 -9 file # same, level 9 is the default'. -s memory now 'at most 2.5 MB' to match bzip2's help (manual table gives 2350k). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This is a very short, unsourced help-wanted stub with misspellings and unconditional compression-efficiency comparisons that are not meaningful without a corpus, level, and implementation. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaces the efficiency-ranking stub with documented block-size controls, tested compression/restore, and integrity/resource guidance. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked bzip2 format capabilities, reference implementations, practical options and resource/integrity boundaries against primary sources. A 37,888-byte mixed binary/text fixture round-tripped exactly with the documented frontend. | 10/8/2026 | 2 |
| Check browser media support | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-haiku-5-5): Tightened the intro and the icon-interpretation section, and removed the closing record-keeping and standards-following paragraphs. The BrowserSupportCheck component import, usage and the VP9 level note were kept. Frontmatter sources and lastVerified are unchanged; description rewritten. Independent review: Checked the component: VP9 probes really use vp09.xx.21.xx at 640x360/30fps/500kbit/s, which exceeds VP9 level 1 limits, so the level 2.1 note is right; fixed its dangling-link sentence. Replaced the generic intro with what the tool actually does. Component and import untouched. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked user-triggered HTML/MSE/MediaCapabilities declarations with explicit geometry and bitrate. No network/media is played and prediction results are not claimed as a successful decode, HDR, DRM or hardware certification. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Checked positive, negative and uncertain declarations against primary browser specifications, separated support from smoothness/power predictions and API limitations, and corrected VP9 codec levels to fit the declared workload. Browser tests cover labels, retry, keyboard use, missing/rejected APIs and absence of media playback. | 10/8/2026 | 7 |
| Choose a codec and encoder | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the generic six-step 'decision order', ffmpeg -encoders/-h preamble and validation boilerplate with concrete recommendations (AV1 10-bit for efficient libraries/delivery, H.264 8-bit 4:2:0 for universal playback, HEVC for TV/HDR, Opus/AAC/FLAC, FFV1 for archival) and the compatibility limits that matter: Safari AV1 only with hardware decode (iPhone 15 Pro/M3+), Android AV1 decoder since 10 and mandatory from 14, H.264 High 10/4:2:2 lacking hardware support, lossless only if the pixel format is preserved, and which codecs MP4/WebM officially carry. Independent review: Verified Android AV1: decoder since Android 10, encoder and decoder mandatory from Android 14 (developer.android.com supported-formats). Safari AV1 only with hardware decode (Safari 17, iPhone 15 Pro/M3) agrees with /reference/video-codec-support/; HEVC browser support described as hardware-dependent, also consistent. MP4 mappings for AV1/VP9/Opus/FLAC and WebM codec list agree with terminology page and container specs; lossless pixel-format caveat correct. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The constraint-first selection process is accurate and useful, avoids claiming that one codec is universally best, and explicitly requires encoder, container, stream, and target-device verification. | 10/8/2026 | 6 |
| Codec Wiki: One Year Later | current | PassReview basis· gpt-6.1-sol · ultra · Correction check Added context identifying this as the original project’s July 2024 anniversary report, with a link to the current privacy notice. Preserved screenshots and visitor totals remain historical author reports; they were not re-audited or presented as current Encode Wiki measurements. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh This is clearly a dated first-person project retrospective rather than current technical guidance; its repository history and locally preserved analytics screenshots support the timeline and reported first-year activity, while speculative audience estimates are labeled as such. | 10/4/2026 | 1 |
| Codecs | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-haiku-5-5): Replaced the 'migrated collection will cover…' placeholder with a two-sentence intro that says what a codec page describes and how it relates to the tool pages. No technical claims were added. Independent review: Two-sentence intro is accurate and plain; no technical claims to verify. | Not recorded | 0 |
| Color Formats | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Tightened the migrated page, fixed the truncated description, restored the original's practical subsampling advice (4:2:0 for most delivery, 4:4:4 for screen recordings/UI) and added 4:2:2 context and chroma location. Corrected the formats table (gbrp vs packed rgb24, added RGBS/gbrpf32le) and removed 'check with -h encoder' preambles and hedging. Independent review: Verified FFmpeg pixel-format names, gbrp vs packed rgb24, yuvj full-range alias behaviour (H.264 full range still decodes as yuvj420p in 9.0.2); no changes needed. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Corrected reversed endian examples, sample-versus-pixel bit counts, planar GBR versus packed RGB mapping, grayscale naming, and duplicate endian labels. Replaced the unsupported dismissal of 4:2:2 with guidance about source and workflow requirements. | 10/8/2026 | 3 |
| Color Primaries | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Rewrote with the H.273 code table using FFmpeg names and real usage, a chromaticity table for BT.709, BT.601-525, BT.2020 and P3 (needed for mastering metadata), and practical notes: mpv's guessing for untagged SD, BT.2020 as the container for P3-graded HDR (tag bt2020, not smpte432), BT.2020 does not imply HDR, Display P3 tagging, and why retagging cannot change gamut. Independent review: Verified code table, chromaticities (BT.709, SMPTE-C, BT.2020, P3, DCI white 0.314/0.351), SVT-AV1 spellings (bt601, smpte240) and mpv's resolution-based guess. Only the x265 source URL was changed to the 4.3-tagged docs. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Separated primaries and white point from transfer and HDR. Clarified DCI-P3 versus P3-D65 and distinguished a primaries identifier from the complete Display-P3 encoding. Unspecified primaries are no longer treated as verified BT.709. | 10/8/2026 | 7 |
| Color Range | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Cut the 'reproducible ramp experiment' and testing narrative; kept a values table and added mix-up symptoms, a signalstats check, retag (setparams, video_full_range_flag bitstream filters, mkvpropedit for the Matroska container value) vs convert (scale/zscale, VapourSynth range_s), the yuvj420p trap, why not to use -color_range, which standalone encoders ignore Y4M XCOLORRANGE (x265 4.3, SvtAv1EncApp 4.2), and Matroska's different Range numbering. Restored the original's point that delivery video including HDR Blu-ray is limited range. Independent review: Wrong: suggested remuxing with mkvmerge to fix the stale Matroska range after video_full_range_flag retagging; mkvmerge keeps the old value. Stream-copy command now adds -color_range tv (tested: container and bitstream agree), mkvpropedit color-range kept for existing files. Added that zscale=range=limited on yuvj420p input converts samples but leaves the output tagged pc (tested); caution box scoped to re-encoding. Verified: signalstats YMIN/YMAX check, scale=out_range=limited,format=yuv420p (0-255 -> 16-235, tagged tv), XCOLORRANGE handling per encoder, SvtAv1EncApp --color-range full / x265 --range full, Matroska Range numbering (1 broadcast, 2 full) and mkvmerge --color-range TID:n. Earlier pass: 2026-09-08GPT-6 Astra · not recorded Corrects automatic-clamping implications, distinguishes nominal code values, metadata repair and pixel conversion, and adds a tested ramp: full luma 0–255 becomes 16–235 through scaling while tag-only changes preserve samples. HDR and unknown source interpretation remain explicitly outside the recipe. | 10/8/2026 | 5 |
| Colorimetry: inspect before encoding | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Rewrote as the colorimetry hub: one table of the five properties with ffprobe field names and typical HD SDR/HDR10 values, a tested ffprobe command with sample output, how to find per-frame HDR side data, mpv 0.41's actual guessing rules for untagged video (restored from the original import's mpv heuristics, re-verified), and a short preserve/retag/convert section. Removed the 'this October 2026 update…' narrative and hedges. Independent review: Side-data hint changed from 'Dolby Vision RPU Data' to 'Dolby Vision Metadata' (present for both HEVC and AV1). Verified mpv 0.41.0 (latest) guessing rules in video/csputils.c and video/mp_image.c: matrix BT.709 if w>=1280 or h>576; primaries from matrix (2020/709) else by height 576 -> BT.601-625, 480/486 -> BT.601-525, else BT.709; transfer BT.1886 for limited range, sRGB otherwise. (The original import's '488' was wrong; 486 is correct.) Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Replaced the stub with an ffprobe inspection workflow and a table separating pixel format, matrix, primaries, transfer, range and chroma location. Distinguished preservation, retagging and conversion. The exact inspection syntax passed on a synthetic 10-bit FFV1 input. | 10/8/2026 | 4 |
| Compare encoding quality with metrics and visual checks | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Rebuilt around how the community compares encodes: one lossless reference with shared preprocessing, several quality levels with rate-quality curves and BD-rate (bjontegaard package), scoring without film-grain synthesis (-export_side_data film_grain), alignment checks, SSIMULACRA2/Butteraugli/XPSNR/CVVDP as primary metrics with VMAF NEG optional and PSNR/SSIM as sanity checks, summarising with mean + low percentile + worst frames, and inspecting the worst sections at normal speed and frame by frame. Removed the PSNR/SSIM/VMAF-centric workflow with settb/setpts/shortest/-v info/-n padding and the testing narrative; one short reproducibility note kept. Independent review: Method matches community practice (one lossless reference, several quality levels, BD-rate, grain handled, alignment check, mean + low percentile + worst frames, visual inspection); kept the structure. Film-grain advice was incomplete: exporting grain only helps if the encoder didn't denoise. Added that SVT-AV1 --film-grain-denoise 1 or aomenc --denoise-noise-level still leave a clean picture scored against a grainy reference (SVT-AV1 4.2.0 default denoise is 0). Added the colour-tag trap with FFVship --verbose (measured: untagged pair -25.7 vs -14.0 across Vship front ends; tagged pair identical). Clarified 'compare at equal size, not equal CRF'; VMAF row now 'NEG or v1' since v1 has NEG built in. Verified bjontegaard 1.3.0 bd_rate(rate_anchor, dist_anchor, rate_test, dist_test, method) signature and sign; tested with plain and negated Butteraugli inputs. Tested -export_side_data film_grain on an SVT-AV1 film-grain encode (PSNR rises without grain), the FFVship JSON jq query, and xpsnr reference-first. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The PSNR, SSIM, and libvmaf examples use the documented distorted/reference input order, while the surrounding guidance correctly stresses alignment, per-frame data, reproducibility, metric limitations, and visual confirmation. Earlier pass: 2026-09-08GPT-6 Astra · not recorded Builds a common FFV1 reference and two x264 candidates, requires timeline alignment, names the VMAF model and retains per-frame evidence; exercises 240-frame PSNR/SSIM/VMAF logs and identity checks without turning synthetic command tests into a quality benchmark. | 10/8/2026 | 8 |
| Condor | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page pins the current condor branch commit, accurately describes its workspace packages and development status, and avoids presenting it as stable Av1an. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): The page pinned a July 2026 commit, Rust 1.95, a non-existent 'condor-site' workspace member and said there were no builds. Rewritten: what Condor is (Andean Condor library + California Condor TUI), its pipeline and metrics, unreleased status, rolling binaries and Docker image from the BoatsMcGee fork, the bundled-model layout and libffms2 requirement, the README's non-working pacman/cargo install lines, and tested basic and target-quality commands. Independent review: Removed default-valued options from the examples (--encoder svt-av1 and --target-metric ssimulacra2 are the defaults per condor --help) and stated both defaults once. Verified subcommands, 0.0.1-unstable version and option names with the rolling build (rev f60af14). | 10/8/2026 | 3 |
| Containers overview | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced the 'compatibility is a relationship' framework and verify-list with a practical chooser (MKV/MP4/WebM/MOV/Ogg/AVI) and remux guidance, including the fact that FFmpeg's default stream selection silently drops subtitles it cannot copy. Independent review: Verified remux claims with FFmpeg 9.0.2: with default stream selection an SRT track is silently dropped going into MP4; with -map 0 the same remux fails ('codec not currently supported in container'). Page wording is correct. MP4 row said 'Only a few subtitle formats (mov_text)'; changed to 'Text subtitles only as mov_text'. Quoted appliesTo FFmpeg version for consistency. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The overview accurately treats compatibility as a relationship among container variant, codec mapping, metadata, muxer, and player, and its comparison and validation checklist are useful and well qualified. | 10/8/2026 | 3 |
| Contributing after the Codec Wiki migration | current | PassReview basis· gpt-6.1-sol · ultra · Correction check Removed the obsolete standing refresh-branch target and directed contributors to task-specific checkout instructions; ordinary integration uses master. Preserved attribution and the existing review status. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The contribution workflow is materially obsolete: it names the old Codec Wiki, clones a different GitHub repository, requires Node 18 and Yarn/Docusaurus, and directs contributors through main and deployment branches instead of the current GitGud Starlight project workflow. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaces obsolete repository/Node18/Yarn/Docusaurus/direct-deploy instructions with the repository's actual pnpm/Starlight review workflow, retaining original provenance link. | 9/7/2026 | 2 |
| CVVDP | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Removed duplicate H1/unused import, the long fcvvdp-derived algorithm walkthrough with an fcvvdp-only short-name display table, and the 'not installed or executed' narrative. Now covers JOD meaning (10 = no difference, 1 JOD ~ 75% preference), when to use it (temporal, colour, HDR, display-specific), pip install and the cvvdp CLI with the options that matter, the upstream standard display models, HDR via standard_hdr_pq/hlg, the resolution/full-screen-resize pitfall, Vship (FFVship and VapourSynth with cumulative _CVVDP) and fcvvdp, plus a five-bullet 'how it works'. Independent review: standard_phone row said only 'Phone display'; filled in from display_models.json (6-inch 2400x1080, 500 cd/m2, 250 lux, 0.4 m). Verified against v0.5.7 (latest on PyPI and GitHub, 2026-08-02): JOD 10 = no difference, can drop below 0, 1 JOD = 75% preference; display model specs; --full-screen-resize behaviour; heatmap modes; config-paths; Vship display names and running _CVVDP score; DOI 10.1145/3658144 resolves to the ACM TOG 2024 paper; reference-package score changes across releases (0.4.0 recalibration, 0.5.3 fix, 0.5.7 temp-padding default) support the 'note the version' advice. Tested: FFVship -m CVVDP --displayModel standard_fhd; JSON is a running score whose last value equals the printed video score. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added first-party ColorVideoVDP 0.5.7 CLI usage and actual display-model names. Distinguished the attributed fcvvdp process overview from first-party defaults, corrected sRGB and PQ/HLG transfer and clipping claims, and removed unsupported score bands and universal speed or equivalence claims. Checked the display-model implementation and distinguished separate tone mapping from transfer decoding. The Python/GPU metric was not executed. | 10/8/2026 | 7 |
| Data compression and archive tools | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced the capabilities table and 'verification boundary' prose with a which-tool-for-what chooser, a strengths/weaknesses table, HTTP coding notes and test commands. Independent review: 'Good default' row used zstd -19 (slow, contradicts the zstd page's 3–9 sweet spot); now 'zstd -T0 file'. Removed -T0 from xz and replaced 'use -T0 with zstd and xz' with the correct per-tool statement. zstd browser support updated to include Safari 26.3 and Edge; sources added. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked compressor versus archive capabilities, standard HTTP content-coding scope, memory/resource limits and exact local synthetic round-trip evidence. Unavailable 7zz/zpaq interfaces remain source-checked. | 10/8/2026 | 6 |
| dav1d | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked 1.5.4 tagged CLI and library defaults separately, including strictness and film-grain behavior for checksum sinks. FFmpeg libdav1d full AV1 decode passed; standalone CLI unavailable. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): The page claimed no standalone dav1d binary was available and was mostly an option dump with limitation caveats. Rewritten: what dav1d is and where it is used, build steps, the fact that the CLI reads only IVF/Annex B/Section 5 (not MKV/MP4, and not stdin) with tested extraction commands, decode/benchmark/MD5 commands, a short options table, and FFmpeg usage. Corrected the FFmpeg film-grain advice: -filmgrain is deprecated in FFmpeg 9; -export_side_data film_grain is the current way to decode without applying grain. Independent review: Film-grain/MD5 claim was wrong as written: `dav1d -o decoded.md5` (muxer picked from the extension) still applies grain; only an explicit `--muxer md5`/`xxh3` or `--verify` disables it (tested: different hashes). Changed the example to `--muxer md5` and rewrote the table row. Verified dav1d 1.5.4 is the latest tag, CLI options/defaults from --help, FFmpeg libdav1d -filmgrain flagged deprecated and -export_side_data film_grain gives the no-grain decode (hash identical to -filmgrain 0), mkvextract/ffmpeg -c copy IVF extraction. | 10/8/2026 | 4 |
| DEAD / distributed-encode | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The pinned commit, early-alpha qualification, patched-Av1an dependency, Redis/RQ architecture, audio and VapourSynth limitations, and operational warnings match the project documentation. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Corrected the claim that VapourSynth input is unsupported (master added --vs-input and --vs-scene-input in May 2025, although the project's QUICKSTART FAQ still says no) and described jobs accurately (groups of scenes, mkvmerge assembly). Removed the architecture diagram, operational-caution boilerplate and caution box padding; kept requirements (Redis, patched Av1an-verbosity fork, no venv), audio dropping, and the worker/client commands. Independent review: Commands were wrong: there is no distributed_encode.py at the repo root (upstream docs are stale); the shim is src/distributed_encode.py, or the `dead` console script when installed. Fixed both commands and said so. Verified all CLI flags against src/distributed_encode/options.py (incl. --vs-input, --vs-scene-input, --password); Redis default 127.0.0.1:6379; last commit 2025-05-30. | 10/8/2026 | 3 |
| Deband | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced a threshold-by-threshold FFmpeg lecture with the actual workflow. Established by test that FFmpeg's deband (and gradfun) do essentially nothing on 8-bit frames unless converted to 10-bit first, and documented that; added vsdeband f3k_deband (now vszip-based), placebo_deband, deband_detail_mask masking and Grainer, plus the real option meanings from the FFmpeg docs. Independent review: Flagged claim was false: on a synthetic 1-code-step 8-bit staircase, 8-bit `deband` reduced 8x8 block error ~29% and `gradfun` ~55-65% (it dithers); 10-bit `deband` ~39%. Neither is a no-op. Rewrote: deband can only scatter pixels across step edges at 8-bit, 10-bit gives it in-between values; gradfun is 8-bit only (FFmpeg auto-converts down) and FFmpeg's docs say it is for playback, not before lossy compression. Found that `format=yuv420p10le,deband` + libsvtav1/libx264 without -pix_fmt silently encodes 8-bit (FFmpeg negotiates deband's output back down); kept -pix_fmt yuv420p10le and explained why. deband_detail_mask polarity was inverted on the page (mask is white on edges/texture, which MaskedMerge keeps from the original); fixed. f3k_deband thr described as neo_f3kdb scale; range explained as random up to N px. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is an empty Help Wanted stub and provides no debanding guidance or supported implementation details. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the empty stub with a ten-bit FFmpeg deband example, normalized per-plane threshold and search-range explanations, texture-loss cautions, and HDR-specific evaluation. | 10/8/2026 | 4 |
| Decombing | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Was a copy of the deinterlace page. Now covers leftover combing: where it comes from (hybrid sources, failed field matching, interlaced content encoded progressive), how to find it, and repair of only the combed frames with fieldmatch+yadif deint=interlaced, vsdeinterlace.vfm postprocess (with cthresh/mi/chroma meaning) and vinverse. Independent review: VFM chroma advice reversed to match VIVTC docs: disable chroma only when chroma problems (rainbowing) cause false detections. Confirmed vinverse UnboundLocalError with amnt=None in vs-jetpack 2.2.6 (undershoot/overshoot only defined when amnt is set), so amnt=255 workaround is valid. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is an empty Help Wanted stub and does not explain combing detection, field order, or any usable workflow. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the empty stub with idet-based diagnosis, verified field-order selection, a field-rate bwdif example, and guidance separating selective residual-combing repair from telecine recovery. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. | 10/8/2026 | 3 |
| Deep Dive into SVT-AV1's Evolution (Part 1): Presets Analysis from v2.0 to v3.0 | current | PassReview basis· GPT-6 Astra · medium · Correction check Preserves historical results but explicitly defines custom metrics as unvalidated exploratory measurements, fixes luma/chroma and timing-versus-BD-rate claims, explains harmonic zero/negative failures and modified baseline, and scopes preset conclusions. This is not validation of the experiment. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The central comparisons rely on unvalidated custom metrics: VMAF is independently run on chroma planes and arbitrarily weighted, XPSNR source code is modified, Butteraugli's intensity is changed, and SSIMULACRA2 is harmonically pooled; the article explicitly says proving these transformations is out of scope and asks readers to trust them, then draws preset conclusions from them. | 9/7/2026 | 6 |
| Deep Dive into SVT-AV1's Evolution (Part 2): Encoder Parameters Revisited | current | PassReview basis· GPT-6 Astra · medium · Correction check Preserves all experiments/media and reports inherited custom-scale/BD-rate limitations; narrows varboost, tune, bias, sharpness, temporal-filter, fast-decode and tile conclusions, and scopes historical lossless mismatches to the tested pipeline. No subjective validation or rerun is claimed. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh Part 2 inherits Part 1's explicitly unvalidated W-VMAF, W-XPSNR, modified Butteraugli, and harmonic-SSIMULACRA2 methodology and uses those results for prescriptive varboost, tune, sharpness, temporal-filter, fast-decode, and tiling guidance; visual examples do not validate the custom objective scales or their aggregation. | 9/7/2026 | 6 |
| Dehalo | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Was a stub around one dehalo_alpha call that implied darkstr had to be raised. Now explains halos vs ringing, when to fix upstream, a table of vsdehalo functions (dehalo_alpha, fine_dehalo, fine_dehalo2, edge_cleaner, vine_dehalo, hq_dering, smooth_dering), real parameter defaults (brightstr=1.0, darkstr=0.0, lowsens/highsens 50, ss 1.5, luma only), and how to view the fine_dehalo mask. Independent review: Dropped default-valued darkstr=0.0 from the dehalo_alpha example; all stated defaults (brightstr 1.0, darkstr 0.0, lowsens/highsens 50, ss 1.5, rx 2, planes 0) and fine_dehalo.masks.MAIN verified in source. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is an empty Help Wanted stub and offers no definition, filter options, or artifact-risk guidance. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the empty stub with halo diagnosis and a source-pinned vs-jetpack dehalo_alpha example; identified the deprecated vs-dehalo package, explained bright/dark strengths, and disclosed that plugin execution was not tested. | 10/8/2026 | 3 |
| Deinterlace | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Was a single bwdif command with parity flags. Now distinguishes true interlaced, telecined and field-matched progressive sources, shows idet usage, single vs double rate, FFmpeg bwdif (default send_field = double rate) / yadif (default send_frame), QTGMC via vsdeinterlace.QTempGaussMC (havsfunc no longer ships QTGMC), and vsaa NNEDI3/BWDIF bobbers. Independent review: bwdif default send_field (double rate) and yadif default send_frame confirmed with ffmpeg -h; yadif has 4 modes vs bwdif's 2, so 'same options' became 'similar options'. QTempGaussMC().deinterlace(clip, tff) and motion_blur(fps_divisor=2) signatures verified; havsfunc r34 release notes confirm removal; parameter-mapping gist exists. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is an empty Help Wanted stub and provides no field-order, cadence, or deinterlacer guidance. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the empty stub with a bwdif workflow explaining field parity, send_field versus send_frame cadence, filter ordering, and progressive output/audio-sync validation. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. | 10/8/2026 | 5 |
| Denoise | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote around what matters: when to denoise vs keep/synthesise grain, FFmpeg hqdn3d (documented 4/3/6/4.5 defaults, side effects) and nlmeans (s/p/r table, speed warning, restored original example settings), and VapourSynth DFTTest and mc_degrain with a difference-clip check. Removed the 'fixed fps claims' prose and metrics boilerplate. Independent review: hqdn3d defaults (4, then 3/6/4.5 derived; hqdn3d=8 == 8:6:12:9) confirmed against docs. Removed 'does not parallelise well' (nlmeans has slice threading) and the unsourced 's above 10 rarely sensible'. mc_degrain example had thsad=400, which is the default; removed it and documented tr=1/thsad=400 defaults. DFTTest/mc_degrain/MakeDiff difference view tested on a 16-bit clip. Coordinator: fixed a link to the renamed banding section. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The nlmeans parameter table incorrectly labels p as research size instead of patch size, while performance and quality recommendations are presented without a reproducible basis. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected nlmeans p to patch size and r to research/search window; removed unsupported fps and universal quality recommendations and added explicit hqdn3d strength ordering and comparison commands. | 10/8/2026 | 4 |
| Discord | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the 'compatibility sample / reproducible playback checks' page with practical guidance: current official upload limits (free 20 MB since August 2026, Nitro Basic 50 MB, Nitro up to 1 GB, boost levels 50/100 MB), format choice (H.264/AAC MP4 safest; AV1 better at small sizes but older Apple devices only decode it in hardware; HEVC needs -tag:v hvc1 and is unreliable), the bitrate budget formula with a worked example, and a tested two-pass x264 command. Restored the original community hvc1 tip. Independent review: Verified via Discord's Help Center API (article updated 2026-10-07): free limit 20 MB since August 2026 (up from 10 MB), Nitro Basic 50 MB, Nitro up to 1 GB, boost level 2/3 50/100 MB, size now checked before compression on mobile, MP4 with H.264/HEVC/AV1 listed. Removed the unsourced WebM playback claim (Discord's FAQ lists only MP4 codecs). Two-pass example: added -pix_fmt yuv420p to pass 1 so both passes use identical video settings; tested (output 639 vs 628 kbit/s target). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The entry is explicitly incomplete and its fast-changing playback matrix is based on undocumented community testing with no versions, test files, dates, client builds, or primary evidence. Earlier pass: 2026-09-07GPT-6 Astra · medium Removed undocumented playback matrix and replaced it with official attachment scope, a concrete compatibility sample, size budgeting and dated client-validation requirements. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. | 10/8/2026 | 3 |
| Dolby audio codecs and Atmos | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Removed legal/licence note, "checked on date" narrative and bare-codec tables. Added what each codec is, that Atmos is metadata on E-AC-3 JOC/TrueHD/AC-4, a clear FFmpeg capability table (ac3/eac3 up to 5.1, truehd experimental needs -strict -2, no AC-4/Atmos, Atmos objects dropped on decode) and a copy-do-not-re-encode recommendation. Independent review: Atmos decode was described as 'channel bed only, objects dropped'. Non-Atmos decoders (incl. FFmpeg) play the backward-compatible 5.1/7.1 mix rendered from the full mix; only the object data and height positioning are lost. Rewrote the TrueHD/E-AC-3 JOC bullets and the FFmpeg table row; merged the redundant 'convert Atmos' row. Blu-ray AC-3 'core' rephrased: it is an AC-3 stream interleaved in the TrueHD PID (FFmpeg mpegts.c exposes it as a sub-stream); dropped the MakeMKV claim. Tested: ac3/eac3 640k from 5.1(side) work; a 7.1 input to eac3 is silently downmixed to 5.1 (now stated); truehd needs -strict -2 and is 5.1-max (mlpenc.c CODEC_CH_LAYOUTS) and bit-exact (MD5 match). E-AC-3 32 kbit/s-6 Mbit/s and 7.1; TrueHD 44.1-192 kHz, 24-bit, 16 channels, Blu-ray 8 ch/96 kHz and 6 ch/192 kHz all confirmed on Dolby pages; ffprobe Atmos profile strings confirmed in profiles.c. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page usefully and accurately separates AC-3, E-AC-3, TrueHD, AC-4, and Atmos, while qualifying proprietary authoring, carriage, decoder, and licensing boundaries. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Extends decoder and browser guidance while distinguishing AC-3, E-AC-3, TrueHD, AC-4, Atmos metadata and rendering. MIME/configuration queries are explicitly conditional rather than claimed device-lab results; encoding ordinary E-AC-3 does not imply Atmos authoring. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Repaired the native TrueHD command's missing experimental opt-in, documented the reviewed implementation's rates/layouts and narrow successful PCM hash check, and removed a duplicate body H1 because Starlight supplies the frontmatter title. No Atmos/Blu-ray/multichannel certification inferred. | 10/8/2026 | 10 |
| dovi_tool | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The 2.3.3 release, Rust baseline, available binaries, command groups, conversion modes, examples, and Dolby Vision workflow boundaries agree with official documentation. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked 2.3.4 tagged command scopes, mode/drop/crop limits and metadata frame alignment. Examples are source-checked; no licensed Dolby Vision sample or final display validation was performed. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Replaced the reviewed-scope table, license notes, validation checklist and a misleading section ('no licensed sample executed') with a workflow-first page: install, a five-step re-encode workflow (extract-rpu incl. -m 2 for profile 7 and FFmpeg pipe, info, x265 --dolby-vision-profile/--dolby-vision-rpu with the required VBV, inject-rpu, mux with mkvmerge or MP4 -strict unofficial), a verified mode table, a command table and editor usage. All commands were run against the 2.3.4 release binary. Independent review: Removed the unnecessary -bsf:v hevc_mp4toannexb from the FFmpeg pipe (FFmpeg's raw HEVC muxer inserts it; extracted RPU byte-identical with and without, including from an MP4 with audio). Added what happens without -strict unofficial in MP4 (no DOVI box, plain HDR10). Re-ran with the 2.3.4 release binary (latest): generate, info --summary/--frame, extract-rpu (MKV, -m 2, pipe), inject-rpu, export -d all/scenes/level5, editor remove/duplicate, plot -p l2, convert/remove on MKV (prints 'Matroska input is experimental!'); x265 4.3 --dolby-vision-profile 8.1 --dolby-vision-rpu succeeds with VBV and fails without; mkvmerge 102.0 writes the DOVI record. Mode table, Rust 1.88 minimum, fontconfig/internal-font and release asset list match the README/release. | 10/8/2026 | 5 |
| DSV | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Cut the maintenance and sample-format hedging down to what the codec is, its features and limits, and the two reference commands. Confirmed the repository was archived on 2024-10-04 and that its README points to DSV2. Fixed the decode example to read the encoder's output file, and noted the -nameValue (no equals sign) option syntax. Independent review: Built DSV1 and ran the page's encode/decode commands successfully. Confirmed the repo is archived (2024-10-04) and that the feature list matches the README. Added the author's MPEG-1/2 efficiency comparison, the header-only dsv1.h, and -fmt for non-4:2:0 raw input (from the encoder's help). Earlier pass: 2026-10-04gpt-6.1-sol · ultra Verified the DSV1 archive date and successor against the upstream repository. Documented 8-bit/project-specific format limits, archived reference decode workflow and unsupported general deployment assumptions while retaining the original examples. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Replaced truncated description and universal comparative/implementation claims; identified DSV1_spec_r2.pdf and single-header reference form, preserved archived syntax, and made developer compression claims attribution explicit. | 10/8/2026 | 1 |
| DSV2 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Removed the wavelet-transform tutorial, the CineForm formula tangent and the validation prose. Kept features, an up-to-date build line (sources are in src/, not the root), and encode/decode usage with the real gotchas: -qp is a quality percentage with 100 lossless, -y4m=1 is required, and without -nfr the encoder exits 254 at EOF. Minimum frame size is 64 pixels, confirmed against the encoder. Independent review: Built master (124cc71). Confirmed the encode/decode commands and that the encoder exits 254 without -nfr. The README and CHANGELOG confirm the v2.8 freeze on 2026-08-29 and that sources live in src/. Removed a testing-narrative clause about a lossless round trip. Replaced it with the README's qp guide (~70+ visually lossless, 100 mathematically lossless) and the CRF default of 85 from vhelp. Cut the 'at least 64 pixels' minimum, which conflicts with vhelp's stated minimum of 16 and is not documented upstream. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Corrected the v2.8 freeze date to August 29, 2026 from the current upstream README, checked multi-chroma/8-bit restrictions and reversed quality-scale direction with qp=100 lossless. Distinguished lossless input samples from forced output conversion and unverified browser deployment. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Replaced truncated description and unsupported generic CineForm/wavelet essay with source-specific subband/quantization discussion; scoped developer comparison, pinned 124cc711 bitstream 2.8/encoder 15/decoder 3, tested src/*.c build, qp 60/100 round trips and exact lossless hashes, documented stale PDFs/build README, actual even and >=64 dimensions, explicit Y4M/stdin/stdout behavior, EOF exit 254 and bounded -nfr automation. | 10/8/2026 | 2 |
| eac3to | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Removed 'this review could not access' narration and hedges; explained what eac3to is for, its current maintenance (madshi stopped at 3.36; the DG/rationalqm.us developers publish newer builds, with regular 3.x releases through 2026), Blu-ray title ')' vs track ':' syntax from the usage guide, common jobs (-core, delays, -slowdown) and an option table from the author's help. Independent review: Cut 'regular releases through 2026' (only builds up to 3.60 in a March 2026 snapshot could be confirmed). -core also extracts the AC3 core of TrueHD; -demux extracts the main video and all audio tracks (not subtitles); -slowdown/-speedup reworded to the documented 25 <-> 23.976 conversion. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Replaced unsupported abandonment and latest-release claims with the original author’s 3.36 help and an explicit successor-forum access limit. Added input listing, selected-track FLAC usage, documented controls, and container and object-metadata cautions. Windows and licensed decoder execution were unavailable. | 10/8/2026 | 3 |
| ECM | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote around the plain facts: ECM is JVET's post-VVC research software on top of VTM, not an HHI codec and not a standard. The original's claim that it 'will likely become H.267' is gone. Kept the build and an encode/decode example from ECM-20.0 (BSD-3-Clause, tag dated June 2026) and removed the research-hygiene and no-guarantee prose. Independent review: Replaced the unsourced 'informal name H.267' sentence with the sourced process. ITU-T and ISO/IEC issued a joint Call for Proposals beyond VVC in July 2026 (JVET-43). Responses are evaluated January 2027 and the first version is targeted for end of 2029 (MPEG 155 press release, ITU JVET page). Bit-depth note was wrong for the shipped config: encoder_randomaccess_ecm.cfg sets InternalBitDepth 10, so ECM codes at 10-bit even for 8-bit input. The page now says so and notes decoder -d (OutputBitDepth) and the rec.yuv the config writes. Confirmed from the GitLab tag API that ECM-20.0 (2026-06-23) is the latest tag. Checked every encoder and decoder option name against the ECM-20.0 EncAppCfg.cpp and DecAppCfg.cpp, and GOP/IntraPeriod 32 and QP 32 in the RA cfg. Cut an unverified 'common test conditions use 22/27/32/37' line that I had drafted. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh It misleadingly presents ECM as a Fraunhofer-developed codec and speculates that it will likely become H.267; ECM is a joint JVET exploration/reference model, not a finalized successor codec or assigned Recommendation. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Verified ECM as revision-specific JVET exploration software rather than a finalized standard or stable production profile. Added source-revision/configuration, depth/chroma, matching-decoder and browser/container evidence boundaries; did not invent version-specific commands from a limited repository web view. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Replaced the repository-access gap with ECM-20.0 tag/commit identity and source-checked 64-bit CMake build, exact RA configuration/encoder/decoder paths, depth/chroma/frame/QP parameter scopes/defaults and native output-depth behavior. Explicitly states commands were not built/timed locally and do not establish deployable standard profiles. | 10/8/2026 | 7 |
| Embedding the Un-Embeddable | current | PassReview basis· gpt-6.1-sol · ultra · Correction check Identified the Discord observations and helper-service list as historical material from October 2023. Distinguished Open Graph video metadata from actual platform fetching and playback support, and linked the current Discord guide. Current helper services and account limits were not retested. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The article consistently frames its service list and reverse-engineered Discord behavior as an October 2023 snapshot, accurately explains the Open Graph video-tag mechanism, and separates observed limitations from assumptions while including conspicuous legal and ethical cautions. | 10/4/2026 | 2 |
| Encode Wiki | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Prose-only edit: plain description and tagline (dropped 'workflows you can verify'), clearer card descriptions, and a two-sentence project note without the 'will be expanded and reviewed' filler. Components, hero actions and PageDataLinks unchanged. Independent review: Prose-only homepage: card descriptions and project note are plain and accurate; all card hrefs resolve (/articles/ is a separate content collection). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Checked both homepage Ask links against the deployed protected team-preview route and verified the sign-in description; the production Ask hostname currently redirects elsewhere, so the working preview is linked. Earlier pass: 2026-10-05T02:43:58.154Zgpt-6.1-sol · xhigh Updated both homepage Ask links to the production hostname requested by the owner and removed the team-preview sign-in description. Public Wiki navigation, attribution and machine-readable content links are retained; development Ask remains a separate protected hostname. | Not recorded | 0 |
| Encoding Animation with SVT-AV1: A Deep Dive | current | PassReview basis· gpt-6.1-sol · ultra · Correction check Identified the results as the author’s SVT-AV1 1.8.0 experiment on the stated anime corpus and hardware. Qualified universal claims about tiles and deblocking and described chunking as a separate throughput choice. Preserved all charts, media and attribution without rerunning the experiment. Confirmed that the original metric plugin was written in Zig and now carries an archive notice. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh For its explicitly scoped SVT-AV1 1.8.0 animation context, the article identifies the exact encoder build, hardware, five short samples, CRF span, and single-metric limitation; it labels its parameter findings early, calls for visual confirmation, and presents them as a starting point rather than current-version truth. | 10/4/2026 | 3 |
| Encoding basics | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the truncated description and Codec Wiki mission text (which also mislinked 'discard data' to psychovisual and 'lossy' to video-artifacts) with a clear format/encoder/decoder distinction, a worked uncompressed-bitrate figure (~600 Mbit/s for 1080p24 8-bit 4:2:0), codec examples per media type, and containers/remuxing/transcoding. Kept the #what-is-a-codec anchor used by other pages. Independent review: Checked the uncompressed-bitrate figure (1920×1080×1.5×8×24 ≈ 597 Mbit/s), format/encoder/decoder distinction and #what-is-a-codec anchor kept. | 10/8/2026 | 1 |
| Encoding guides | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the meta text about tested versions, validation steps, the content-review report and 'an AI pass is not human approval' with a reading order for the seven guides plus a short table of recipes that live on codec/tool pages. Independent review: Reading order and recipe table checked; every linked route exists. | 10/8/2026 | 0 |
| Eve-AV1 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the 'parameter coverage gap' tables and generic ffprobe/ffmpeg check commands with a short factual page: Two Orioles' commercial AV1 encoder (EVE family, started as VP9), the vendor's 20% / 3–5× claims attributed and qualified (no methodology, no independent tests), licence and evaluation only, standard AV1 output, open alternatives. Independent review: Checked against twoorioles.com: 20% bitrate reduction or 3–5× speed claim, 'Netflix (720) - June 2025' chart label, VoD focus, subscription service and evaluation on request, and dav1d co-developed with VideoLabs and MultiCoreWare (as Two Orioles states it). No changes needed. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Replaces unqualified vendor performance claims with a source-checked evaluation boundary and an explicit unavailable full-parameter/default inventory. No proprietary flags, defaults or licensed execution are invented; standard output-validation commands are clearly not Eve encoding syntax. | 10/8/2026 | 2 |
| Eve-VP9 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Removed the coverage-gap table, the ffprobe/ffmpeg check commands and the old vendor VMAF chart. Now a short factual page: Two Orioles' commercial VP9 encoder (first in the EVE family), the vendor's compression claims and table totals (−22.26% vs libvpx 1.15.1, −38.45% vs x264, −13.82% vs x265 4.1) attributed and qualified, licence and evaluation only, standard VP9 output, link to vpxenc. Independent review: The vendor table lists 19 clips, not 20 (counted in the page HTML); corrected. Percentages (−22.26% vs libvpx 1.15.1, −38.45% vs x264, −13.82% vs x265 4.1) and 'nearly 20%/40%' text match. Earlier pass: 2026-10-04Codex (runtime model identifier not recorded) · not recorded Checked the public product offering, speed/thread descriptions and vendor chart attribution. No release-specific CLI/API or defaults are asserted; implementation/profile/browser boundaries are explicit. The probe and full-decode commands passed on a synthetic libvpx VP9 fixture, not a licensed Eve output. Independent review found no remaining factual issue. | 10/8/2026 | 2 |
| exhale | current | PassReview basis· claude-opus-5-5 · high · Correction check Preset-rate table matches exhaleApp.cpp at v1.2.2; replaced 'no fixed rate claimed' for expert presets 10–12 with the rates exhale reports (208/256/288 kbit/s stereo) and added mono/5.1 scaling; corrected -V/-v descriptions; argument rows rewritten. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Coordinator: input sample-rate limits per preset taken from exhaleApp.cpp (32.1 kHz + 12 kHz per preset step; eSBR needs at least 22.05 kHz). Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks the complete exhale 1.2.2 positional/expert interface and eight executed examples. All encoded/probed as xHE-AAC; four SBR outputs decoded strictly. Non-SBR FFmpeg decoder limitations and success-with-errors behavior are disclosed, with fail-fast verification commands. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Page was a reference with verification narrative and hedges. Rewritten with what exhale is (only free xHE-AAC encoder; 1.2.2 is the latest and ends active development), presets verified against source (0-9 = 16n+48, a-g = 12n+36 kbit/s stereo; letter modes use eSBR, mono/stereo only, 22 kHz+), expert mode, stdin use, and the real FFmpeg decoding limitation; argument table kept verbatim. Independent review: Low-preset warning named presets '0 and a, 36 kbit/s and below', but preset 0 is 48 kbit/s; the README's warning is about the lowest mode only (18 mono/36 stereo = preset a). Fixed. 1.2.2 (tag 2025-12-20) and 'end of active development' confirmed in Release.htm; preset formulas 16n+48 and 12n+36 (eSBR) and the eSBR limits (>=22 kHz, mono/stereo) confirmed in exhaleApp.cpp; numeric presets reject >=7 channels. Re-tested with J1's exhale build: FFmpeg 9.0.2 decodes preset b (eSBR) but fails presets 5 and 9 with 'Not yet implemented in FFmpeg' while exiting 0, so the -xerror advice is right. | 10/8/2026 | 4 |
| FAQ | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-haiku-5-5): Removed the unverifiable list of contribution requirements (attribution, primary evidence, command versions, uncertain-claim marking) and the 'approval does not establish' and 'deployment and promotion are separate' filler; kept the start-here, provenance, benchmark-reading and resources answers in plain prose. Verified the Codec Wiki repository is live and not archived; lastVerified updated to 2026-10-08 and sources limited to what was accessed. Independent review: Rewrote robotic answers ('Read an article's versions ... as part of its results'); restored the original page's reason for the wiki's existence (no cohesive entry point), and replaced the vague contribution answer with the actual workflow verified in the repo's CONTRIBUTING.md (merge request to master on gitgud.io, primary sources, tested versions). Renamed the benchmark question heading (no inbound anchors). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Replaced stale claims about other guides with current beginner links, codec and implementation distinctions, review-status guidance, and limits on interpreting dated benchmarks. Preserved the original Codec Wiki migration history. | 10/8/2026 | 2 |
| Fast Computation of SSIMULACRA2 on GPUs: A Performance Evaluation | current | PassReview basis· GPT-6 Astra · medium · Correction check Preserves timing data and preliminary scoring findings while withdrawing TurboMetrics/stride quality recommendations; explains unvalidated implementation differences, sampling bias and differing decoder workloads, and corrects video-threshold/GOP/NVDEC generalizations. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The timing experiment is useful, but the article recommends TurboMetrics for quality assessment with stride 3 or 5 after its own preliminary validation found GPU scores statistically different from the reference and no linear correlation; calling those unretuned scores an alternative metric does not establish perceptual validity. | 9/7/2026 | 5 |
| FFMetrics | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The v1.7.0 Windows scope, metric set, CLI shape, model/build caveats, reproducibility checklist, and unresolved upstream license status are accurately and usefully documented. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Updated to current upstream (stable 1.7.0, beta 1.8.0-beta.2) and made consistent with the metric pages: says plainly it only offers FFmpeg's PSNR/SSIM/VMAF/XPSNR and points to Vship for SSIMULACRA2/Butteraugli/CVVDP. Added the important caveat that 1.7.0 passes XPSNR inputs in the wrong order (fixed in 1.8.0 beta 1), the colour-range caveat from the README, real install steps and the documented CLI options. Removed the PowerShell 'environment check', reproducibility checklist and licence essay. Independent review: Wrong CLI option: -vmaf-upscale-to-model was renamed -vmaf-scale-to-model in 1.7.0 beta 8 (confirmed in changes.txt shipped in the 1.7.0 zip and in FFMetrics.exe strings); the GitHub README is stale. Fixed the table row. Corrected the scaling description: since 1.7.0 beta 4 it scales up or down to the model's resolution when inputs differ by more than 10%, otherwise scales the encode to the reference. Example now uses XPSNR (with the 1.8.0 beta) and NEG VMAF instead of SSIM, in line with the metric pages. Verified XPSNR input-order bug fixed in 1.8.0 beta 1 ('Swapped ref & main in XPSNR template'), .NET Framework 4.8 (exe.config supportedRuntime), up to 24 files, bundled vmaf-models. | 10/8/2026 | 7 |
| FFmpeg | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The current 9.0.1 release, build-dependent feature model, core tools, command examples, option placement, probing, validation, and licensing caveats align with official FFmpeg documentation. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked FFmpeg 9.0.2 input/output stream option scopes, explicit mapping, probe fields and full-decode boundaries; exercised synthetic AVC/FLAC encode, JSON probe and complete decode. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted two optional FFmpeg -map values so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Removed the 'reviewed release' table, -version/-buildconf/-encoders 'basic verification' ritual, reproducibility checklist and long parameter dump. Wrote a concise orientation: what FFmpeg is, where to get good builds per OS (gyan.dev full vs essentials verified, BtbN, Homebrew/evermeet, distro lag), licensing in one paragraph, positional options, automatic stream selection and -map with a table of tested specifiers, -c copy with the MP4/ASS failure case and keyframe-cut behaviour, filters and filter_complex, piping 10-bit Y4M to standalone encoders (verified -strict -1 is still required in 9.0.2), ffprobe and a decode check, and documentation links. Independent review: Verified gyan.dev essentials vs full (svtav1/dav1d/rav1e only in full), `winget install ffmpeg` = full build, latest 9.0.2; -strict -1 still required for 10-bit yuv4mpegpipe in 9.0.2 and the SvtAv1EncApp -i - pipe produces 10-bit. | 10/8/2026 | 7 |
| FFV1 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced the preamble-heavy command and hash-validation narrative with one archival command and an option table, explaining -g 1 once. Fixed option facts from the 9.0.2 source: -coder 1 and 2 both select the custom-table range coder, slice CRCs are on by default in version 3, version 4 is experimental and its IETF draft has expired, and bit depths run 8 to 16. Restored the version history from RFC 9043 and the original's intra-only explanation. Independent review: Wrong default slice count: the draft said FFmpeg picks 4 for HD. ffv1enc.c (n9.0.2) sizes slices to about 360x288 or smaller, which gives 20 at 1080p. Confirmed by matching output sizes with and without -slices 20. The row is rewritten. -level 3 and -slicecrc 1 are defaults since FFmpeg 8.0 (n7.1 vs n8.0 source). The page now says so and gives the real reason to pin them: older builds could pick version 0/1 for SD. Dropped -context 1 from the recommended command. On real 10-bit film it made files about 0.5% larger and on 8-bit anime about 1% smaller, so it stays as a table row described as content-dependent. Kept -coder 1, which was 6% smaller on 8-bit; FFmpeg forces the range coder above 8 bits. Added the archival practicalities: -map 0 to keep all tracks, and -dn because Matroska rejects MOV tmcd tracks (tested; the timecode survives as a tag). The lossless check now writes framemd5 files and diffs them (tested identical). Fixed the RGB caveat: hashing the source as bgr0 does not match, but decoding the FFV1 file with -pix_fmt gbrp does. Version history wording matches RFC 9043 (stable 2006/2009/2013, v2 experimental only). Confirmed the v4 draft (-22) has expired and that v2/v4 need experimental strictness in ffv1enc.c. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page mixes accurate RFC 9043 material with speculative v4 claims and an unsupported absolute archival rule that GOP size must be 1; current guidance needs to distinguish inter-frame context dependence, keyframes, slice CRC behavior, and an archive's actual policy. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked RFC 9043 sample representation, extra-plane, coder and version distinctions, and FFmpeg implementation limits. Explained FFV1 level as a bitstream version, preservation validation, and separate web access copies without asserting universal hardware/native-browser support. Locally smoke-tested FFmpeg 9.0.2 version-3 output and confirmed exact 10-bit 4:2:2 source-to-decoded frame hashes. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added runnable SHA256 decoded-sample validation for an actual 10-bit 4:2:2 source, confirmed exact local FFmpeg 9.0.2 version3 round trip, separated aggregate samples from timing/metadata/file integrity, and quoted the optional audio map for zsh. | 10/8/2026 | 4 |
| FLAC | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Removed the licence admonition, "validation and reproducibility" checklist, hash/version narration and duplicated format tables. Now: when to use FLAC, flac and FFmpeg commands, what compression levels do (size/time only), metaflac, subset note, compatibility. Independent review: FFmpeg -compression_level 11 and 12 set max LPC order 32 (flacenc.c); tested on a 44.1 kHz signal, flac -a showed orders 23-32, i.e. outside the streamable subset (max order 12 at <=48 kHz). The page claimed FFmpeg defaults stay in the subset and 9-12 are just 'slower and slightly smaller'; now warns that 11-12 leave the subset and to stay at <=10 for hardware. Subset description now mentions LPC order. flac 1.5.0 confirmed as latest tag; flac -8 -V, flac -t and metaflac tag/picture commands tested. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The format limits, lossless semantics, metadata, container distinctions, CLI workflow, and validation advice align with RFC 9639 and FLAC 1.5.0 documentation. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds RFC 9639 full-format rate, precision and channel limits, streamable-subset boundaries, reference/native implementation roles and container-specific browser cautions. A flac 1.5.0 verify/test plus full FFmpeg decode preserves the tested 32-bit mono PCM exactly. | 10/8/2026 | 5 |
| GIF | current | PassReview basis· claude-opus-5-5 · high · Correction check Gifsicle 1.96 and FFmpeg 9.0.2 encoder/muxer/palettegen/paletteuse snapshots verified current; every option, alias and default present. Rewrote ~100 descriptions (-gifimage, palettegen max_colors/reserve_transparent, the 1.96 --resize-touch-height width bug). Independent review: No wrong rows in about 25 sampled. Gifsicle: bare --lossy 20, bare --threads 8, loopcount -2, conserve_memory 0, ignore-errors on by default, optimize levels. Every FFmpeg n9.0.2 gif encoder, muxer, palettegen and paletteuse default and range matches its AVOption table. Coordinator: the appendix now links to the consolidated 'Reading option tables' page. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The checklist is too thin and incorrectly says progressive decode is unsupported; GIF89a defines interlaced image data specifically to support progressive display, and transparency is palette-index based rather than a general alpha channel. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks all Gifsicle 1.96 canonical controls/aliases and FFmpeg 9.0.2 GIF encoder, muxer and palette filters. Corrects omission-versus-bare values and misleading gamma help, labels resize-touch-height dispatch failure, and limits preservation claims to fully decoded frames/timing and the sample-matching optimized fixture. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds GIF89a palette, alpha, timing, interlacing and metadata boundaries, including local per-frame palettes and lossless indices versus lossy quantization. Adds encoder/decode/compositing guidance while preserving the pinned full Gifsicle and FFmpeg parameter inventories. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): About 3,300 words of Gifsicle/FFmpeg option inventories and a 'Validation and limits' section replaced by a practical guide: why GIF is poor for video-like content, FFmpeg palettegen/paletteuse as a single filtergraph and a two-step command with the options that matter (stats_mode, dither, diff_mode), gifski (defaults 90 quality, ~800x600 width cap, fps only for PNG input), Gifsicle -O3/--lossy, and pitfalls. The five bound tables are kept as an appendix. Independent review: gifski command was wrong in practice: it applied fps=15 in FFmpeg and claimed --fps only matters for PNG input. gifski 1.34.0's CLI resamples video/Y4M input to --fps (default 20) by dropping/duplicating frames, so the pipe produced 20 fps with duplicated frames. Now pipes plain Y4M and sets --fps 15 --width 640 in gifski, with the explanation corrected. Tested both FFmpeg palettegen/paletteuse forms (single split filtergraph and two-step palette.png) with FFmpeg 9.0.2; outputs identical and correct size/frame count. Option table (stats_mode, new=1, bayer dither, diff_mode, max_colors, -loop) matches FFmpeg docs. Replaced the generic 'tool versions this page was written against' appendix intro with the actual tool versions. Coordinator: linked the option appendix to the image encoder parameter reference. | 10/8/2026 | 10 |
| Graining | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Was a stub. Now separates keeping grain, adding pixel grain (vsdeband.Grainer, FFmpeg noise with correct t/u flag meanings) and AV1 film grain synthesis (SVT-AV1 v4.2.0 --film-grain/--film-grain-denoise/--fgs-table/--adaptive-film-grain, aomenc denoise/table options), with decoder-side and metric consequences. Independent review: Grain-synthesis description was wrong for SVT-AV1: with the default --film-grain-denoise 0 it encodes the original source and only adds the grain model; aomenc encodes the denoised frames by default (--enable-dnl-denoising=1). Rewrote the intro and table row. Option names/defaults checked with SvtAv1EncApp v4.2.0 and aomenc --help; Grainer protect_edges meaning checked in source. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is an empty Help Wanted stub and contains no grain-generation, synthesis, or validation guidance. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the empty stub with a seeded FFmpeg noise example and explanations distinguishing retained grain, added pixel noise, and AV1 decoder synthesis, including denoise and decoder-setting checks. | 10/8/2026 | 3 |
| gzip | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Condensed to gzip levels/-n/--rsyncable, pigz usage (pigz 2.8 built and tested), multi-member files and HTTP notes. Independent review: Verified gzip -k/-9/-t/-l, -n reproducibility (identical bytes with -n, different without), concatenated members, --rsyncable. Only prettier reformatting. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The guide is incomplete and incorrectly describes GNU gzip's -a option as the way to compress multiple files; -a is the legacy ASCII/text-mode option, while multiple file operands are handled independently without it. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrects -a misuse and single-file/archive distinctions, adds tested GNU gzip round-trip and tar workflow, and explains size fields and reproducible headers. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked gzip format capabilities, reference implementations, practical options and resource/integrity boundaries against primary sources. A 37,888-byte mixed binary/text fixture round-tripped exactly with the documented frontend. | 10/8/2026 | 3 |
| HDR formats and tone mapping | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced a hedged comparison table, a 'what it cannot establish' evidence table and a testing narrative with a practical HDR page: format table (HDR10/HDR10+/Dolby Vision/HLG), how to identify each from ffprobe output, HDR10 static metadata explained, HDR10+ carriage, a Dolby Vision profile table (5, 7, 8.1, 8.2, 8.4, 10) with the RPU re-encode workflow, HLG, and tested libplacebo and zscale+tonemap HDR-to-SDR commands. Added the verified pitfall that FFmpeg 9 carries HDR10 SEI into a tone-mapped SDR output unless removed with sidedata=mode=delete. Dropped the RelationshipTable import; HDR10+ Advanced, Dolby Vision 2 and SL-HDR are reduced to one verified sentence. Independent review: Identification table said Dolby Vision frames show 'Dolby Vision RPU Data'; that only appears for HEVC. AV1 profile 10 frames (libsvtav1/libaom-av1 -dolbyvision 1) show only 'Dolby Vision Metadata', which HEVC also shows, so the table now uses that. Removed the four-step Dolby Vision re-encode list and the HDR10+ encoding details that duplicated the guide and dovi_tool page; replaced with one paragraph each linking to the guide (the main how-to) and the tool pages. Verified: zscale+tonemap chain runs; libplacebo option names (colorspace, color_primaries, color_trc, range, format, apply_dolbyvision default true) against FFmpeg 9.0.2 -h filter=libplacebo; the libplacebo command runs (lavapipe needs -init_hw_device vulkan on this host, real GPUs don't); both chains leak HDR10 MDCV/CLL into the SDR output as SEI and as Matroska elements unless sidedata=mode=delete:type=MASTERING_DISPLAY_METADATA / type=CONTENT_LIGHT_LEVEL is added (type names checked against -h filter=sidedata). Verified the profile table against FFmpeg n9.0.2 dovi_rpuenc.c (5 → compat 0; 8 → compat 1/2/4; AV1 → 10; profile 7 EL encoding unsupported) and the x265 4.3 docs (profiles 5, 8.1, 8.2, 8.4); Dolby Vision 2 date (2 Sept 2025) confirmed via search.
Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately separates HDR transfer systems, static/dynamic metadata, codec/container mappings, and tone mapping, and carefully dates emerging HDR10+ Advanced and Dolby Vision 2 claims. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added current in-force BT.2100-3 identification and practical stream/frame inspection with a field interpretation table. Separated sample precision, chroma and metadata from faithful HDR presentation; documented software tonemap linear-float prerequisites without publishing a universal conversion recipe. Synthetic HEVC mastering/content-light SEI inspection passed, with explicit visual/dynamic/display exclusions. | 10/8/2026 | 11 |
| hdr10plus_tool | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The 1.7.2 release, Rust baseline, supported inputs, commands and option ordering, examples, and metadata-order caveat match the official README. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked 1.7.2 command scopes, ignored verify during injection/removal and inclusive plot endpoints. No mastered HDR10+ sample, final display behavior or end-to-end injection is claimed. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Removed the duplicated scope/validation/parameter sections and testing narrative; restored the original import's practical extract/inject/remove commands (dropping its obsolete -strict -2 and -vbsf), and added the full re-encode path (extract -> x265 --dhdr10-info or -x265-params dhdr10-info, or inject), --verify, the editor JSON, plot options and the README's wrong-order (--skip-reorder) diagnosis. All commands run against the 1.7.2 release binary. Independent review: Pipe command trimmed to 'ffmpeg -i input.mp4 -c:v copy -f hevc -' (removed -map 0:v:0 and -bsf:v hevc_mp4toannexb; output JSON identical, audio is not selected for the raw HEVC muxer). Added that x265 only honours --dhdr10-info when built with ENABLE_HDR10_PLUS, which defaults to OFF in x265 4.3's CMakeLists. Re-ran with the 1.7.2 release binary (latest): extract (HEVC, MKV, pipe), --verify extract, inject, remove, editor, plot; x265 --dhdr10-info with hdr10plus_tool's test JSON; README claims (Rust 1.85, formerly hdr10plus_parser, Samsung scene info, no -o = verify only, --skip-reorder diagnosis) checked. | 10/8/2026 | 4 |
| HEIC | current | PassReview basis· claude-opus-5-5 · high · Correction check Snapshots pinned libheif 1.23.5; moved heif-enc, x265 and Kvazaar snapshots, heading and tables to 1.23.6 (f81f28ac). 1.23.6 removed the registered-but-ignored --even-size option, so the row and snapshot entry were deleted; plugin sources are unchanged between tags, help/source hashes recomputed, and the encode/decode recipes re-run on the 1.23.6 binary. Rewrote all descriptions (x265 complexity now documents its real thresholds: rd 6/rd-refine at 60, cu-lossless at 70, WPP off at 90); replaced the numbered 'Parameters 1–20' summaries with first-to-last option names. Independent review: --chroma-downsampling default changed to 'average' in this pass; heif_encoding_options_alloc sets average as preferred with only_use_preferred=false, so it is not forced. Default now 'average (preferred, not forced)'; description says -C forces the choice. --bit-depth claimed 'HEVC tops out at 15'; the x265 plugin codes 8, 10 or 12 bits, and -b also applies to TIFF input. Rewritten. --add-pyramid-group was described as grouping 'tiled images'; the source groups several input images into a multi-resolution pyramid. Fixed. --add-pyramid-group, --set-gimi-track-id, --sai-data-file and --enable-metadata-compression said 'hidden from --help'; they are always registered and are listed only in experimental or header-compression builds. Reworded. --tiling-method called unci experimental; unci is gated by WITH_UNCOMPRESSED_CODEC and only tili by experimental features. Fixed. --hevc default 'default codec' was vague; now 'false (HEVC is the fallback)', and the description states the suffix fallback rule. --raw-type did not say it is required for raw input, and --raw-width/--raw-height did not say at least one is needed (decoder_raw.cc). Added. Checked and correct: --even-size removal, quality/thumb/lossless fallback, fps 29.97 and 90000-tick rounding, repetitions/max-frames rejecting 0, NCLX defaults 1/13/6/1, unci zlib, gop low-delay, OUTPUT derivation and every x265/Kvazaar row (CRF=(100-q)/2, complexity thresholds 60/70/90). Coordinator: the appendix now links to the consolidated 'Reading option tables' page. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page conflates HEIC, HEIF, and a codec, makes unsupported camera-implementation and Android-transcoding generalizations, and reduces current HEIF/HEVC precision and lossless capabilities to an inaccurate fixed checklist. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Reviews libheif 1.23.5 frontend registrations, hidden/build-gated flags and fixed x265/Kvazaar plugin contracts, with the verified x265 4.3 parser at its current upstream host. Standalone x265 passthrough defaults are not conflated with plugin defaults; Kvazaar is source-only. Executed recipes separate reversible RGB/alpha, presentation rotation, grids and sequence edit-list repetitions from untested device/HDR interoperability. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds HEIF item capabilities and HEVC Main/Main 10/range-extension chroma and precision boundaries. Separates libheif parser/plugin availability, auxiliary images, video decoding, Safari HEIC display and fallback delivery; preserves prior complete frontend/plugin references. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): The page was ~3,500 words, mostly a capability table and 'Worked HEIC parameter combinations' with test narratives, and it lacked the practical point that HEIC is a capture/interchange format to convert for the web. Rewritten with what HEIF/HEIC/HEVC profiles mean, patent/licensing status, platform support (Safari 17+, Android 9+ decode, Windows Store extensions, no Chrome/Edge/Firefox per caniuse), converting with heif-dec/heif-enc, an options table, and pitfalls (grids, metadata, gain maps, 4:2:0). The bound heif-enc frontend, x265 and Kvazaar tables are kept verbatim at the end. Independent review: Verified heif-enc/heif-dec options against libheif 1.23.6 help: x265 plugin defaults preset=slow, tune=ssim, chroma=420; -b applies only to 16-bit PNG input (default 10); heif-dec output suffixes and -q for JPEG. Added the tune default to the table. Android row verified against the Android supported-formats page (decode 9+, device optional, 10-bit needs Main10; encode 10+ via MediaCodec/HeifWriter). Safari 17 HEIC support confirmed in WebKit's Safari 17 notes. Rewrote the banned 'Treat HEIC as…' sentence, fixed a capitalisation typo, and replaced the unsourced 'Platform vendors pay' with the OS-codec vs bundled-codec distinction. Ran the page's heif-enc/heif-dec/heif-info commands and -p chroma=444 locally. Coordinator: linked the option appendix to the image encoder parameter reference. | 10/8/2026 | 8 |
| HEVC / H.265 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Rewrote with when-to-use guidance, one profile table including range-extension profiles, tier/level facts (High tier from level 4; common level examples), plain-language coding tools (incl. the community practice of reducing SAO), encoders, hardware/browser summary, containers with the hvc1/hev1 Apple pitfall and a tested retag command, HDR formats with real x265 option names, and updated licensing: Access Advance acquired Via LA's HEVC/VVC program (now VCL Advance) in Dec 2025, and Velos Media's joint pool ended. Removed the duplicated workflow checklist and capability-boundary tables. Independent review: Confirmed Access Advance acquired Via LA's HEVC/VVC program effective 15 Dec 2025, now 'Video Codec Licensing Advance (VCL Advance)'. Velos Media was presented as a current separate licensor; its joint pool wound down (Qualcomm and the other licensors left). Rewrote to say Velos stopped joint licensing and holders license individually. Tested: libx265 to MP4 writes hev1; -c copy -tag:v hvc1 retags. Tested ffprobe -show_frames -read_intervals %+#1 shows mastering-display and CLL side data on an x265 HDR10 encode. Verified MV-HEVC decoding in FFmpeg 7.1 changelog, current ITU-T edition 01/2026, High tier from level 4, level 4.1/5.1 sample-rate limits, SVT-HEVC repository archived. Coordinator: HDR10 mastering-display and content-light SEI are per frame, so the check now uses ffprobe -show_frames -read_intervals %+#1 (verified on an x265 HDR10 encode).
Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page explicitly calls H.265 version 10 from July 2024 the in-force edition, but version 11 from January 2026 is now in force and version 10 is superseded. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected the in-force H.265 edition to version 11, January 2026. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked Main/Main10 versus Range Extension depth/chroma scope, x265 input/internal precision and profile restrictions, software/hardware distinctions and final-file/browser validation. Preserved existing practical guidance and separated HDR precision from metadata and rendering. | 10/8/2026 | 8 |
| High dynamic range | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Removed hedges and the 'no HDR display was tested' note; now states plainly what HDR is (transfer function, not bit depth), PQ (absolute, to 10,000 cd/m²) vs HLG (relative, gamma + log segments, BBC/NHK), a format table (HDR10 static ST 2086 + MaxCLL/MaxFALL, HDR10+ ST 2094-40, Dolby Vision RPU and profile fallbacks 5/7/8.1/8.4), and what an encode must keep, including that FFmpeg forwards mastering/CLL side data to libx265 and libsvtav1 while dynamic metadata must be re-injected with hdr10plus_tool/dovi_tool. Independent review: Cut the duplicated workflow section (static/dynamic metadata handling, encoder option names, FFmpeg side-data passthrough, tone-mapping filters) to two short paragraphs linking to /guides/preserve-color-and-hdr/ and /hdr/. Dolby Vision profile fallback statements (8.1→HDR10, 8.4→HLG, profile 5 no fallback, profile 7 dual layer) could not be checked against a Dolby primary source (Dolby pages return 403); cut them and linked /hdr/#dolby-vision, which owns that table. Kept HDR10+→HDR10 fallback. Dropped the SVT-AV1 Parameters source the page no longer uses; description updated. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Corrected HLG’s square-root/logarithmic scene encoding and separated gamut, transfer and dynamic range. Clarified HDR10+ metadata interchange and Dolby Vision’s profile-dependent layers. Added cautions about dynamic-metadata frame alignment and the difference between metadata edits and color transforms. | 10/8/2026 | 5 |
| HM | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page makes unsupported universal quality claims against x265 and tells readers to use an option whose meaning the page admits it does not know. Earlier pass: 2026-09-07GPT-6 Astra · medium Removed universal quality superiority claims over x265 and the unexplained -xPS option; pinned HM-18.0 with a Main10 raw-input configuration, QP/GOP explanations, and encoder-versus-decoder reconstruction validation. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Reduced to what HM is, why it is not a practical encoder (no threading, fixed intra period, raw YUV only), build for HM-18.0 (latest tag, April 2023) and one encode/decode example. Dropped the unverified claim from the original that HM beats tuned x265 on quality, the 'did not compile' narrative and the reproducibility checklist; the mysterious -xPS flag from the original is not used. Independent review: Built HM-18.0: the documented build fails on GCC 16.2 because HM forces -Werror (unused-but-set-variable, array-bounds); -Wno-error is overridden. Added the tested -DCMAKE_CXX_FLAGS=-w workaround. Ran the page's encode/decode exactly (10-bit raw, randomaccess_main10.cfg, -wdt/-hgt/-fr/-f/--InputBitDepth=10/-q 32): Main 10 stream, decoder output md5-identical to recon; binaries in bin/ as stated; doc/software-manual.pdf exists; cfg names exist. Verified no Y4M support, no thread option, IntraPeriod only (no scene-cut) in TAppEncCfg.cpp; HM-18.0 (2023-04-20) is the latest tag. Commands now use ./bin/ paths (binaries are not on PATH). | 10/8/2026 | 4 |
| Image and audio decoder support | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-haiku-5-5): Removed the 'source-checked on 2026-10-04' framing, the 'Reproducible validation record' checklist and the 'detection patterns, not results from a lab' paragraph, and rewrote the JPEG XL section. The Chrome 155 status is now stated from the sources: Google's announcement says Chrome 155 decodes JPEG XL, Chrome Status lists stable as scheduled for 6 October 2026, and the rollout is staged. Frontmatter sources and lastVerified unchanged; description rewritten. Independent review: JPEG 2000 (coordinator/R5b note) corrected: WebKit's Safari 18.0 notes say 'WebKit for Safari 18.0 removes support for the JPEG2000 image format' and that Safari was the only browser to ever support it; row now says no current browser decodes it, and the Safari 18.0 post is cited. JPEG XL statuses verified: Chrome 155 stable shipped 2026-10-06 (chromiumdash) with jxl-rs; Firefox 158 release notes list JPEG XL and the release calendar gives 2026-10-13; MDN BCD notes Safari 17 does not support animated JPEG XL. Removed outdated 'experimental flag' wording. Safari 18.4 Ogg Opus/Vorbis platform list verified. Tightened the intro and removed a closing 'use the vendor announcement and test' hedge and 'does not establish' phrasing. Coordinator: aligned JPEG XL browser support with Chrome Status (155), MDN BCD and the Firefox 158 beta notes. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Source-checks dated browser/OS/carriage matrices and Chrome 155 beta versus its future stable schedule. Provides guarded image/MIME/WebCodecs detection patterns and reproducible validation boundaries; explicitly reports source verification rather than a comprehensive browser/device test. | 10/8/2026 | 11 |
| Image codecs overview | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Coordinator: links now point to the consolidated 'Reading option tables' page instead of the retired parameter pages. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The comparison correctly separates coding, file structure, metadata, and application support and gives a useful, qualified overview without pretending format support is universal. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Adds the shared parameter-guide entry point and JPEG 2000/QOI mappings, corrects the dead WebP specification source, and distinguishes ordinary DCT JPEG from predictive lossless modes without asserting target application support. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds all nine image sample/chroma/alpha/animation comparisons and links dated support checks. Distinguishes theoretical format limits from application builds, includes specialty JPEG/JPEG 2000/QOI boundaries, and treats Chrome 155 default JPEG XL as incoming after the verification date. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Aligned the PNG/APNG overview row with the W3C third edition: APNG is included in that specification and animation still needs distinct feature checks. Other codec capability and dated browser claims retained. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): The page was a dense capability matrix with defensive caveats and a 'choosing a page shape' note about the wiki itself. Replaced with a 'which format should I use' table (web photos, quality/HDR, existing JPEGs, graphics, transparency, archives, animation, QOI), a compact format comparison with browser support (Chrome 155 JPEG XL default-on, Safari JPEG XL/JPEG 2000) and four short cross-format tips. Independent review: JPEG 2000 browser support said 'Safari only'; WebKit removed JPEG 2000 in Safari 18.0. Changed to 'None (Safari dropped it in 18.0)' and added the WebKit source. HEIC browser column was vague ('Apple platforms; elsewhere needs OS or licensed HEVC decoders'); now 'Safari 17+ only' (MDN/caniuse/WebKit 17 notes). JPEG XL support listed 'Firefox 158+' as if shipped; Firefox 158 stable is due 2026-10-13. Aligned with the coordinator's wording (Safari 17+ stills only, Chrome 155+, Firefox 158 due 13 October). Fixed the confusing 'quality numbers' tip (it listed avifenc and then 'AVIF uses different scales again'); linked Butteraugli; PSNR/SSIM wording matches the brief's consensus. Coordinator: aligned JPEG XL browser support with Chrome Status (155), MDN BCD and the Firefox 158 beta notes. | 10/8/2026 | 9 |
| Interactive media and embeds | current | PassReview basis· claude-opus-5-5 · high · Correction check Documented the compact forge cards, ProjectCard, the committed forge snapshot and pnpm forge:refresh, with examples matching the new components. | 10/8/2026 | 7 |
| Intro | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced a "how to read capability claims" table with a short primer on sample rate, bit depth, channels, bitrate, lossless vs lossy and perceptual coding; removed the review banner and the Nyquist hedging. Independent review: Checked sample-rate/Nyquist, 16-bit ~96 dB, PCM bitrate arithmetic (48 kHz x 24 bit x 2 = 2304 kbit/s), lossless/lossy framing and the ffprobe command; all correct and concise. No edits. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page labels itself incomplete and gives an oversimplified sampling explanation: two samples described as a peak and trough are not sufficient to establish faithful reconstruction, and the theorem requires a band-limited signal plus sampling above the highest occupied frequency's Nyquist rate. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Explains how to evaluate precision, sampling, channels, losslessness and spatial/browser claims. Separates lossy output buffer depth from preserved samples, codec timing from source rate, and audio dynamic range from video HDR/color properties. | 10/8/2026 | 2 |
| Inverse Telecine | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced a single fieldmatch command and generic caveats with how 3:2 pulldown works, hard vs soft telecine, idet detection, the FFmpeg fieldmatch,yadif,decimate chain, VapourSynth VFM/VDecimate (8-bit only, order mandatory) and the vsdeinterlace wrappers/Wobbly. Found that vssource.BestSource applies RFF by default whereas core.bs does not. Independent review: VIVTC restriction corrected: VFM is 8-bit only, VDecimate accepts 10/16-bit (tested on R81). FFmpeg dejudder,fps=30000/1001,fieldmatch,decimate chain and the fieldmatch+yadif=deint=interlaced+decimate chain match the FFmpeg docs; vfm/vdecimate/sivtc/jivtc signatures verified. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is an empty Help Wanted stub and does not explain cadence recovery or provide a usable inverse-telecine workflow. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the empty stub with a verified 3:2 fieldmatch/yadif/decimate workflow, explained recovery from 30000/1001 to 24000/1001 fps, and warned against fixed decimation of native or mixed video. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. | 10/8/2026 | 6 |
| ISO Base Media File Format | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote as a single clear explanation of how ISOBMFF, MP4, MOV, 3GP, HEIF/AVIF and CMAF relate, with a box/sample-entry table; removed the 'what it does not establish' table and the verification paragraph. Independent review: ISOBMFF/MP4/MOV/3GP/HEIF/CMAF relationship is explained once and correctly; MOV brand 'qt ' confirmed with ffprobe on an FFmpeg-written .mov. 'FFmpeg only exposes tracks' was inaccurate (FFmpeg demuxes HEIF items); reworded to 'ffprobe reports streams, not boxes or item properties'. Removed an 'In practice' tic. ffprobe brand command tested. Opus-in-ISOBMFF draft status confirmed (opus-codec.org document is marked incomplete). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately distinguishes ISOBMFF from MP4 and MOV, explains brands and boxes, and correctly describes AV1 mapping while qualifying the incomplete Opus binding and real-player support. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added brands/sample-entry/configuration distinctions, HEIF item-versus-track inspection boundaries, timescales/composition/edit timing, faststart versus fragmented/profile-governed streaming, and a tested selected-track MP4 remux. Dated the Xiph Opus binding's incomplete 0.6.8 status and explicit QuickTime exclusion without denying actual MP4 Opus implementations. | 10/8/2026 | 5 |
| JM | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This is a Help Wanted stub with only a short historical comparison and no version, build, interface, configuration, or validation information. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the historical Help Wanted stub with JM-19.1 CMake build guidance, the supplied QCIF baseline configuration, named input/profile/QP fields, and ldecod reconstruction checks. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Rewritten as a short page: what JM is, build for JM-19.1 (latest tag, April 2023), how the bundled QCIF sample config runs, which config keys to change, how to decode. Removed the provenance/'did not execute' narrative. Independent review: Built JM-19.1 (tree reassembled from jm19.0.zip plus vcgit raw files, hash-verified against the tag). Build fails on GCC 16 (-Werror in lencod/block.c, ldecod); added -DCMAKE_C_FLAGS=-w -DCMAKE_CXX_FLAGS=-w. Binaries are bin/lencod_static and bin/ldecod_static, not 'lencod'/'ldecod' on PATH; fixed the cfg/ example to ../bin/lencod_static and ../bin/ldecod_static (tested: 3-frame Baseline encode, decode OK). Claim that JM lacks rate control was wrong (RateControlEnable, Bitrate in bps); reworded the x264 comparison and mentioned JM rate control. Added QPBSlice and that encoder_main.cfg uses B-frames. Verified config keys, ProfileIDC values, QP 0–51, JM-19.1 (2023-04-19) latest tag, sample clip and configs in cfg/. | 10/8/2026 | 3 |
| JPEG | current | PassReview basis· claude-opus-5-5 · high · Correction check libjpeg-turbo 3.2.0, MozJPEG 4.1.5 and jpegli 031a0077 verified current. Rewrote all descriptions; fixed MozJPEG -quant-table (invalid value is rejected, not ignored), the -restart note, -icc; fixed jpegli --quality (quality 90 maps to distance 1.0, so the default isn't 'not applied'), INPUT (single - reads stdin) and the single-frame claim; corrected the abbreviated commit to 031a0077. Independent review: libjpeg-turbo -optimize said 'always on for progressive and 12-bit'; jcmaster.c also forces it for lossless. Fixed. Changed defaults checked and correct: MozJPEG -dc-scan-opt 0 (jcparam.c, help says 1), -icc (rdpng.c copies iCCP), -tune-hvs-psnr, stdin/stdout; jpegli --quality default (quality 90 maps to distance 1.0). MozJPEG -quant-table rejects out-of-range values; the quality-based sampling thresholds are 80/90. T3b's open issue confirmed: MozJPEG 4.1.5 built from the tag, default progressive with -restart 1 on a new 333x217 fixture gives 'Corrupt JPEG data' under djpeg -strict in both MozJPEG and libjpeg-turbo 3.2.0; -baseline or -revert -progressive decode cleanly. -restart 4B passed on this fixture, so the row's wording ('fails strict decoding') is accurate for row intervals at least. Coordinator: the appendix now links to the consolidated 'Reading option tables' page. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Reviews separate libjpeg-turbo 3.2.0, MozJPEG 4.1.5 and standalone JPEGli source-snapshot inventories and defaults. Corrects help/source default disagreements and quality-scale assumptions, labels predictive lossless versus DCT modes, and documents a reproduced progressive/restart decoding failure instead of recommending that combination. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds baseline/progressive/extended/predictive-lossless/arithmetic coding distinctions, chroma and print modes, and encoder/decoder roles. Preserves libjpeg-turbo, MozJPEG and JPEGli inventories and bounds broad browser compatibility to ordinary delivery modes. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Corrected the overview's one-scan-only baseline claim using T.81's complete-per-component versus progressive refinement distinction. Replaced incorrect variable-frequency-count quantization wording with the fixed 64-entry step-size model. Preserved complete encoder parameters and earlier execution evidence. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): About 5,200 words of parameter inventories, PNM fixture narration and a 'complete versioned' framing replaced by a practical guide: when to use JPEG, libjpeg-turbo vs MozJPEG vs jpegli, starting commands, progressive vs sequential, chroma subsampling pitfalls (MozJPEG and ImageMagick pick 4:4:4 from quality 90), lossless recompression with cjxl/djxl and jpegtran, and a short DCT pipeline explanation restored from the original. The three bound parameter tables are kept verbatim in a 'Full option reference' appendix. Independent review: Verified MozJPEG 4.1.5 set_quality_ratings in rdswitch.c: -quality >= 90 selects 1x1 (4:4:4), >= 80 selects 2x1 (4:2:2), only when -quality is given. The page's MozJPEG example at -quality 80 therefore produced 4:2:2 while presented as a web default; added `-sample 2x2` with a comment. Replaced the unverified 'can nearly double the file' wording. Verified jpegli defaults to 4:4:4 (jpegli_set_defaults sets all sampling factors to 1), -d recommended 0.5–3.0 and -q 68–96 from cjpegli help, and Google's '35 %' high-quality claim from the jpegli blog. Verified locally (libjpeg-turbo 3.2.0): cjpeg reads PNG; -progressive gives the same output as -progressive -optimize; jpegtran's default -copy comments drops the Exif orientation and -copy all keeps it; -rotate 90 -perfect works. Pitfall 'Strip deliberately' (banned word) rewritten as concrete advice (apply rotation and convert to sRGB before stripping). Butteraugli linked; Firefox wording aligned with the 158 schedule. Coordinator: aligned JPEG XL browser support wording with the JPEG XL page. Coordinator: linked the option appendix to the image encoder parameter reference. | 10/8/2026 | 10 |
| JPEG 2000 | current | PassReview basis· claude-opus-5-5 · high · Correction check OpenJPEG 2.5.4 verified current; all 38 rows match opj_compress.c. Rewrote every description, shortened -W/-m default cells; documented that -m is not a valid option, -x is ignored with a warning, POC tile numbers start at 1. Independent review: No wrong rows in about 10 sampled: -GuardBits 0–7, -TargetBitDepth TIFF-only, -mct auto 1 for at least 3 components, -threads only applied when >= 1 (OPJ_NUM_THREADS otherwise), -IMF levels, -n default 6. The changed -m/-W defaults are correct. Coordinator: the appendix now links to the consolidated 'Reading option tables' page. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page incorrectly says DCP image essence is often losslessly compressed; the DCI profile uses the irreversible 9/7 transform and describes its JPEG 2000 delivery compression as visually lossless, not mathematically lossless. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Reviews OpenJPEG 2.5.4 registrations, initial/profile/thread defaults and full decode/sample checks. Labels advertised-but-unregistered -m, broken -x, source-only legacy/JPWL paths and cinema/IMF boundaries; examples distinguish full reversible layers from truncated/irreversible output and DCP authoring. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds JPEG 2000 standard-part, precision, component sampling, wrapper/alpha, scalability, cinema/IMF and HTJ2K distinctions. Preserves the full OpenJPEG reference and avoids interpreting theoretical family maxima, DCP authoring or historic Safari paths as general browser support. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): About 2,800 words of opj_compress inventory and fixture narration replaced by a concise explanation of where JPEG 2000 is still used (DCI cinema verified against the DCI spec: 2048x1080/4096x2160, 12-bit, 250 Mbit/s cap; DICOM; archives; geospatial), the file/variant table, encoder landscape (OpenJPEG 2.5.4, Kakadu, Grok, OpenJPH, FFmpeg) and a few OpenJPEG commands. The opj_compress table is kept verbatim as an appendix. Independent review: Page said JPEG 2000 'never got web support' and that 'Safari is the only major browser engine that decodes JPEG 2000 natively'; Safari supported it until WebKit removed it in Safari 18.0. Both sentences corrected; source added. Confirmed OpenJPEG README wording ('This repository should be considered unmaintained'), latest tag v2.5.4, and Grok README claims (AGPL-3.0, HTJ2K, GDAL integration, grk_transcode, partial decoding). Tested opj_compress/opj_decompress lossless round trip (AE 0), -I -r 20, -r 20,10,1 -p RPCL -threads ALL_CPUS, opj_decompress -r 1 half-size decode, FFmpeg jpeg2000 and libopenjpeg encoders. Rewrote the oddly specific '-m, -x' invalid-option pitfall as a general one: opj_compress warns, ignores an invalid option and encodes anyway (observed locally). Coordinator: linked the option appendix to the image encoder parameter reference. | 10/8/2026 | 10 |
| JPEG XL | current | PassReview basis· claude-opus-5-5 · high · Correction check All 62 cjxl rows checked against cjxl_main.cc and encode.h at v0.12.0 (current); no wrong defaults. Rewrote every description: --epf now explains decoder edge-preserving filter strength 0–3, --streaming_output states what it does, --lossless_jpeg note corrected (lossy -d/-q on JPEG input needs -j 0), added center_x/center_y and effort 11 constraints. Independent review: No wrong rows in about 25 sampled. Confirmed against the source: explicit --intensity_target=0 is rejected (ParseIntensityTarget requires > 0, although help says 0); --upsampling_mode 1 gives pixel dots at 4x/8x (JxlEncoderSetUpsamplingMode); alpha distance 0 when unset; distance defaults; --lossless_jpeg/-d interaction. Replaced the numbered 'Parameters 1–20' style summaries with first-to-last option names, matching heic. Coordinator: the appendix now links to the consolidated 'Reading option tables' page. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately distinguishes the ISO/IEC 18181 format from libjxl v0.12.0 and gives current, appropriately qualified instructions for lossy, pixel-lossless, JPEG-reconstruction, HDR, and progressive workflows. Earlier pass: 2026-09-07GPT-6 Astra · medium Expanded practical guide with version-scoped recipes, controls, tradeoffs, compatibility, and validation. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Reviews all libjxl 0.12.0 cjxl registrations, four-verbosity advanced controls, initialization and input-dependent defaults. Recipes separately verify RGB/alpha/native-depth samples, JPEG-byte reconstruction, progressive completed files and animation timing; direct djxl XMP extraction failure and untested expert/HDR/device paths remain explicit. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds sample/extra-channel/coding-mode and encoder/decoder distinctions and source-verifies Safari, Firefox preview and Chromium release-channel status. Chrome 155 beta decoding and its October 6 stable schedule are distinguished from October 4 stable availability; preserves cjxl references. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): The page was ~4,200 words of capability tables, 'Practical tuning and verification' and 'verified on' narratives, and a browser section that was stale (Chrome 155 'beta'). Rewritten into a practical reference: what JXL is for, lossless JPEG recompression, `cjxl -d`/`-q`/`-e` semantics (distance 1 default, effort 1-10 default 7), JPEG input handling (`-j 0`, `djxl -j`), a task table, and current browser status (Safari 17+ stills only, Chrome 155 stable 2026-10-06 via jxl-rs, Firefox default-on in 158 per beta notes). The bound cjxl 0.12.0 parameter table is kept verbatim at the end. Independent review: Firefox row said released builds ship JPEG XL as an opt-in preview; MDN's 'preview' means Nightly. Now 'Nightly only until Firefox 158, which enables it by default (stable scheduled 2026-10-13)'. JPEG recompression saving '15–20 %' changed to 'about 20 %' with the JPEG XL overview paper (arXiv 2506.05987) as source. Comparison section carried single-photo percentages (20–30 % vs libjpeg, 10–15 % vs WebP) from the author's own test; numbers removed, qualitative comparisons kept, and noted libjxl is tuned with Butteraugli so metric comparisons need visual checks. Starting command dropped `-e 7` (default); 'smaller web delivery' row dropped arbitrary `-e 8`. Noted GIF input defaults to lossless (-d 0) per the bound table. Verified locally (cjxl/djxl 0.12.0): default lossless JPEG transcode with byte-identical djxl restore (cmp), -x strip=exif refusing without --allow_jpeg_reconstruction=0, --progressive_dc, --keep_invisible and --faster_decoding help text and defaults; -q 90 ≈ -d 1 mapping. Coordinator: linked the option appendix to the image encoder parameter reference. | 10/8/2026 | 11 |
| Kvazaar | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Pinned the reviewed v2.3.2 build steps and documented native controls. Distinguished QP 20 from CRF 20, native compile-time depth and P400/P420 input from the FFmpeg 9.0.2 wrapper’s 8-bit yuv420p support. Added raw-input and bitrate-unit guidance and removed a universal x265 ranking. Native building and encoding were not executed. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Had a duplicate H1, a Tabs import, hedges ('a universal quality/speed ranking is not established', 'not executed encodes'), an 'Supported Color Space' table and long tabbed build instructions. Rewritten concisely: Tampere University Ultra Video Group, BSD-3, v2.3.2 (Sept 2025) with active master, uvg266 lineage, positioning vs x265, install (Arch package, MinGW Clang note from README), 10-bit compile flag, FFmpeg wrapper limits (yuv420p only, comma-separated -kvazaar-params), tested commands, and an options table (preset, qp, bitrate in bits/s, vaq, gop, period, open-gop, threads/owf). Added the pitfall that 10-bit builds need --input-bitdepth 10 even for Y4M. Independent review: Verified v2.3.2 (2025-09-17) is still the latest release with master active (pushed 2026-09-30), Arch extra package 2.3.2, README defaults (qp 22, period 64, gop 16 / 8 at fast and faster, open-gop on, vaq 1–20 recommended 5, bitrate in bits/s, lambda RC), MinGW Clang note. Re-ran both commands and confirmed a 10-bit build fails on 10-bit Y4M without --input-bitdepth 10 and produces Main 10 with it. | 10/8/2026 | 5 |
| Lossless compression | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Restructured the broken numbered list (8-space-indented Huffman lines rendered as code blocks) into sections; kept the corrected Huffman and arithmetic-coding examples (re-derived: 1.6 bits, interval [0.43, 0.439), 7/16 decode, entropy ≈ 1.408) without the stacked caveats; added how FLAC, PNG, FFV1 and CABAC/AV1 actually use prediction and entropy coding, dictionary coding, a lossless-formats table and the pixel-format pitfall. Updated NIST URLs to their current host. Independent review: Re-derived the worked examples: Huffman average 1.6 bits, arithmetic intervals [0.43, 0.439), 7/16 decode path, entropy ≈ 1.408 bits. FLAC/PNG/FFV1/CABAC/AV1 entropy-coding descriptions correct. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The arithmetic-coding walkthrough contradicts itself about the fourth symbol, applies the first interval formula incorrectly, repeats new_low where new_high is intended, and labels the resulting interval for inconsistent sequences, making a core teaching example materially wrong. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrects arithmetic coding to one consistent Green/Red/Yellow/Green interval sequence, verifies exact bounds and example fraction, explains termination/model overhead and finite precision, and fixes the RLE tone example. | 10/8/2026 | 5 |
| Lossy compression | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced generic 'choose and validate' advice with the prediction → transform → quantization → entropy-coding pipeline (only quantization loses data), perceptual techniques (chroma subsampling, visual and auditory masking, AV1 grain synthesis, Opus band folding/noise), typical artifacts and generation loss; restored the original's perceptual-coding focus without its incorrect 'removes frequencies outside human hearing' simplification. Independent review: 'Codecs quantize high frequencies more coarsely' is wrong as a general rule (x264, x265 and SVT-AV1 use flat quantization by default); rewrote: small high-frequency coefficients round to zero first, and JPEG/some encoders add quantization matrices. 'Entropy coding … typically with an adaptive arithmetic coder' ignored JPEG/MP3/AAC Huffman coding; now names both variable-length and arithmetic coding with examples. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The opening definition is broadly sound, but the page is an explicit stub that covers only two audio mechanisms, provides no video or image treatment, examples, tradeoffs, or validation guidance, and ends with 'More coming soon,' so it is not yet a useful general introduction. Earlier pass: 2026-09-07GPT-6 Astra · medium Expands the stub into an introduction spanning images, video and audio, separating prediction/transforms from quantization and covering generation loss, rate/effort tradeoffs and practical validation. | 10/8/2026 | 3 |
| Machine-readable access | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Repointed the example section-API links from the removed AV1 'color-and-hdr-signaling' section to the rewritten AV1 page's 'Containers and HDR' section, so the documented example URLs resolve. | 8/22/2026 | 0 |
| Matrix Coefficients | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the bare code list and disclaimers with a full H.273 table (FFmpeg names, KR/KB, where each is used, including ipt-c2 for Dolby Vision profile 5 and YCgCo-R), a section on the BT.601/BT.709 mistake with its three real causes (untagged SD upscales, tagging instead of converting, FFmpeg's default BT.601 untagged RGB->YUV conversion — verified on 9.0.2) and a tested fix command for RGB input, plus the bt2020nc/bt2020c distinction. Independent review: Verified code points and FFmpeg names (incl. ipt-c2 15, ycgco-re 16, ycgco-ro 17) against FFmpeg 9.0.2 setparams/zscale option tables; KR/KB values; SVT-AV1 names (bt601, bt2020-ncl) via SvtAv1EncApp --color-help. Confirmed by test: PNG red -> libx264 with no options gives Y=81 (BT.601), untagged, and 4:4:4; the page's scale=out_color_matrix=bt709,format=yuv420p,setparams command gives Y=62 tagged bt709/bt709/bt709. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Corrected FCC code 4 and distinguished constant-luminance, non-constant-luminance and chromaticity-derived matrices. Added H.273 identifiers 15–17 and explained why resizing or retagging cannot establish a different component transform. | 10/8/2026 | 4 |
| Matroska (MKV/MKA) | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Reframed around what Matroska is for (anything, ASS/PGS, attachments, chapters, HDR metadata) and added tested ffmpeg and mkvmerge/mkvpropedit/mkvextract commands for tracks, languages, attachments, chapters and splitting; removed the test narrative. Independent review: All ffmpeg, mkvmerge, mkvpropedit and mkvextract commands run successfully with FFmpeg 9.0.2 / MKVToolNix 102.0 (track language/name/default flag, attach, OGM chapters, split, extract tracks/attachments/chapters). FFmpeg refuses to write an attachment without a mimetype ('Attachment stream has no mimetype tag'); page said it was so 'players recognise it'. Corrected. 'Browsers and Apple software do not play MKV' was too broad (Chrome plays MKV, Firefox added MKV playback in 2025); changed to Apple software not playing it and browser support being partial. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The structure, metadata, AV1 and Opus mappings, delay/preroll details, and tool distinctions align with RFC 9559 and current Matroska specifications. Earlier pass: 2026-09-07GPT-6 Astra · medium Expanded practical guide with version-scoped recipes, controls, tradeoffs, compatibility, and validation. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted optional video/audio maps for zsh, added FFmpeg derivative overwrite refusal and explained decode-check scope. Existing MKVToolNix 101.0/FFmpeg 9.0.1 examples remain versioned; fresh shell/decode check used FFmpeg 9.0.2. | 10/8/2026 | 6 |
| MediaCodec | current | PassReview basis· claude-opus-5-5 · high · Correction check Replaced range-restating cells with meanings (ndk_codec auto = NDK when no Java VM; pts_as_dts auto when -bf is 0/unset; operating_rate is a speed hint; cq takes -global_quality), corrected the intro to say the other five MediaCodec encoders share every option except -level, and dropped the hedged sentence. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks all 13 AV1 MediaCodec private controls against FFmpeg 9.0.2 source, Android level flags and generic overrides. Removes contradictory unversioned device anecdotes, scopes nominal GOP timing to the wrapper conversion and labels all Android commands source-only. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Removed the 'versioned inventory' and verification narration, the restated Android API caveats and the repeated ffprobe/ffmpeg check commands. Page now says what MediaCodec is, how to find device encoders, gives VBR/CQ examples, explains the real -g behaviour (round(g/fps) seconds), 8-bit-only input, and the level-flag quirk. Option table moved verbatim to 'Full option reference'. Dropped unverifiable Pixel 8-specific claims about broken VP9/AV1 output from the original. Independent review: Added the FFmpeg milestones, verified from allcodecs.c at n6.0/n6.1: H.264/HEVC MediaCodec encoders in 6.0, VP8/VP9/MPEG-4/AV1 in 6.1. -g behaviour corrected: the source does integer division (gop_size / integer fps) before round(), so the remainder is dropped (-g 89 at 30 fps is still 2 s), and FFmpeg's default gop_size of 12 becomes a 1-second interval. The page previously said round(g/fps). 8-bit-only input (yuv420p/nv12/surface), global_quality -> 'quality' key only in CQ mode verified. | 10/8/2026 | 4 |
| MKVToolNix | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The prominent release date is wrong by a year: MKVToolNix v100.0 was released on 2026-07-05, not 2025-07-05, materially misrepresenting how current the reviewed release is. Earlier pass: 2026-09-07GPT-6 Astra · medium Updated reviewed scope to v101.0, released August 24, 2026, including its TrueHD and dependency fixes; corrected the historical v100.0 date to July 5, 2026. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked track IDs, per-input options, extraction and edit selectors; synthetic 102.0 remux/extraction passed. Independent tests retained multiple explicit default audio flags, so the stale automatic-clearing claim was removed. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Removed the release-baseline narration (v101 dependency details, libEBML versions), check-mode/validation checklists and duplicated sections. Restored the original community image and GUI tips (BDMV/mpls import, splitting without re-encoding, WebM option, header editor) and added tested mkvmerge (identify, mux with languages/names/fonts, track selection, split by duration/parts/chapters, append), mkvextract (tracks, attachments, chapters, timestamps) and mkvpropedit examples, plus an option table and the mkvmerge-ID vs mkvpropedit-selector distinction. Independent review: Split behaviour corrected: mkvmerge can only split right before a keyframe, so cuts move to the first keyframe at or after the requested point (not the 'nearest' keyframe), per the mkvmerge man page. Verified 102.0 is current (NEWS 2026-09-14); mux/mkvpropedit examples run. Coordinator: moved the restored GUI screenshot into the GUI section and pointed it at the site's mirrored copies instead of hotlinked GitHub URLs. | 10/8/2026 | 6 |
| MP3 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced generation tables and hedging with LAME -V 0/-V 2 guidance (with typical bitrates), FFmpeg libmp3lame -q:a equivalent, CBR/ABR notes and gapless/VBR pitfalls; removed the -h encoder preamble. Independent review: LAME 4.0 (July 2026) is current; its changelog has no encoder-tuning changes, so the Hydrogenaudio V0 ~245 / V2 ~190 kbit/s figures still apply. 'LAME's default is 4' was misleading: plain `lame` encodes 128 kbit/s CBR (tested: 64 kbit/s for mono input, 'qval=4'); -V defaults to 4 only when VBR is requested. Fixed. Tested ffmpeg libmp3lame -q:a 2 and -b:a 192k -abr 1. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh It cites MPEG-2 audio as ISO/IEC 13818-32 instead of 13818-3 and presents 320 kbit/s transparency as universal rather than content-, encoder-, and listener-dependent. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checks the LAME documented MPEG-1, MPEG-2 and unofficial 2.5 rate/bitrate ranges, mono/stereo boundaries, encoder interface differences and browser/API distinctions. The updated FFmpeg example completed a full synthetic decode; duration/gapless and real-device claims remain bounded. | 10/8/2026 | 5 |
| MP4 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Cut 1200 words of testing narrative and hedging to a practical page: what MP4 holds, -movflags +faststart (box order confirmed), mov_text-only subtitles, hvc1 tagging for Apple, fragmented MP4, chapters showing up as bin_data. Removed '-n', '-map 0:V:0' and ffmpeg -h preambles. Independent review: Codec table listed Vorbis as 'does not fit', but FFmpeg's MP4 muxer writes Vorbis (and PCM) without complaint; Theora, ProRes and DNxHD/DNxHR are the ones actually rejected. Table column renamed 'Avoid or not allowed' and corrected; WebVTT added to unsupported subtitles. SRT handling clarified: default selection silently drops it, -map 0 makes the remux fail; '-map 0:v -map 0:a -map 0:s? -c copy -c:s mov_text' tested (tx3g track). faststart explanation overstated ('must be downloaded almost entirely'); rewritten: players fetch the end of the file first via range requests, or download it all without range support. Verified: default HEVC tag is hev1 and -tag:v hvc1 changes it; chapters written into MP4 show up as a bin_data data stream; +faststart puts moov before mdat; +frag_keyframe+empty_moov+default_base_moof produces moov then moof boxes. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page correctly emphasizes brands, sample entries, fragmentation, streaming profiles, metadata, and target support rather than treating the .mp4 extension as a compatibility guarantee. Earlier pass: 2026-09-08GPT-6 Astra · not recorded Expands codec-versus-container planning, faststart, fragmentation, chapter representation and playback diagnosis; verifies root-level moov-before-mdat order, moof fragments and complete decoding with FFmpeg 9.0.1 while rejecting universal device or CMAF guarantees. | 10/8/2026 | 3 |
| MP4Box | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The GPAC v26.07.0 release, MP4Box scope, build/install boundaries, documented command shapes, and ISOBMFF/DASH validation cautions agree with official project sources. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked official import, inspection, splitting and DASH options, units and RAP boundaries. Binary unavailable; examples remain source-checked and do not claim executed packaging. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Removed 'not installed / documented command shapes' narration, reproducibility checklist and 'format boundaries' hedges. Rewrote around tested tasks: inspect (-info, -diso), mux with -new (and the in-place editing pitfall), import options (:fps=, :lang=, :name=, :dur=, :delay=, #video/#audio selectors), -lang/-rem/-cat/-raw/-split, fragmenting, and DASH and HLS packaging (manifest type follows the -out extension) with keyframe-alignment advice. Independent review: Checked -rap, -lang [tkID=]LAN, -diso, -new semantics against MP4Box 26.07 help (batch M's unpacked binary); -diso writes <name>_info.xml; .m3u8 output produces HLS playlists. | 10/8/2026 | 7 |
| NMKODER | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added the dated 1.10.0 release, Windows/.NET requirements, actual utility workflows and guidance for inspecting generated commands. Distinguished stream copy, subtitle burn-in, metadata tagging and color conversion. GUI and bundled-encoder execution were unavailable. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Removed the 'not executed / not certified' narration and checklist prose; states plainly that it is unmaintained (last release v1.10.0 March 2022, no commits since June 2024), points to maintained alternatives, and keeps the original community tab-by-tab usage plus requirements (Windows 10/11 x64, .NET Framework 4.7.2). Independent review: Verified v1.10.0 (2022-03-28) last release and last commit 2024-06-05. | 10/8/2026 | 2 |
| NVENC | current | PassReview basis· claude-opus-5-5 · high · Correction check The rendered table had no descriptions (type/default/range only). Added a source-checked description for every av1_nvenc option to the snapshot and a Meaning column to the renderer, e.g. legacy slow/medium/fast = p7 two-pass/p4/p1, -rc -1 resolves to constqp with -qp else vbr, -cq needs -b:v 0 because of the 2M generic default, AV1 driver ignores qp_cr_offset, tf_level needs -bf 4, lookahead clipped by -surfaces/-delay. Intro prose and generic-override prose tightened. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh It incorrectly says NVENC encodes VP9; NVIDIA's current desktop Video Codec SDK lists H.264, HEVC, and AV1 for encode, with VP9 in the NVDEC decode list. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected the false VP9 NVENC encoding claim: VP9 belongs to NVDEC decoding, while the SDK encode paths are H.264/HEVC/AV1. Added per-GPU capability discovery and an explicit h264_nvenc preset/tune/VBR example. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks all 45 FFmpeg 9.0.2 AV1 private option spellings, ranges/named values and wrapper defaults, plus codec-specific generic initialization overrides. Hardware recipes are independently source/help-checked, not GPU-executed; driver negotiation, automatic settings and compatibility remain explicit limits. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Page was a generic 'discover the installed path' preamble, an SDR H.264 command at an arbitrary 5M bitrate, the AV1 option inventory first, and 'was not executed' narration. Rewritten as a practical guide: hardware-vs-software trade-off (canonical section), codec/GPU generations (HEVC 10-bit Pascal+, HEVC B-frames newer Turing+, AV1 Ada+), CQ command (-rc vbr -cq -b:v 0, with the wrapper's 2M default bitrate as the reason), AV1 variant, streaming CBR, cuda full-GPU pipeline, options table, pitfalls (-global_quality rejected, driver/nv-codec-headers mismatch). Marked parameter block moved verbatim to 'Full option reference'. Independent review: Myth corrected: the 9.0.2 wrapper (and every release since at least 4.4) zeroes averageBitRate/VBV whenever -cq is non-zero ('CQ mode shall discard avg bitrate'), so '-b:v 0' is not needed with -rc vbr -cq. Removed it from commands; the page now explains that the 2 Mbit/s default only bites when neither -cq nor -b:v is given, and suggests -maxrate for a ceiling. -cq range text fixed: 0 is the default and leaves CQ off (code checks `if (ctx->quality)`), useful values are 1-51 (H.264/HEVC) / 1-63 (AV1). GPU table corrected against NVIDIA's support matrix: HEVC 10-bit exists on Pascal and on the 6th-gen GTX 1630/1650 parts (they only lack HEVC B-frames; the page said they lacked 10-bit too); HEVC from Maxwell 2nd gen; Blackwell adds H.264 10-bit as well as 4:2:2 (SDK 13.0 read-me). Wrong pitfall fixed: unsupported features (B-frames, temporal AQ, lookahead, 10-bit, weighted prediction) make nvenc_check_capabilities return ENOSYS, i.e. FFmpeg stops with '... not supported'; they are not silently ignored with a warning. Session limit made concrete (12 on current GeForce drivers per the matrix; professional cards unrestricted). -highbitdepth now noted for H.264 on Blackwell; uhq described as HEVC/AV1 only; noted p6/p7 may enable lookahead from the preset (nvenc.c comment). Hardware-vs-software section and its anchor kept. | 10/8/2026 | 4 |
| Observing SVT-AV1 v2.1.0's improvements: A New Deep Dive | current | PassReview basis· GPT-6 Astra · medium · Correction check Withdraws universal representativeness and preset/hardware/rate-control recommendations unsupported by five short anime clips; preserves all experiments/graphs, states missing visual validation and historical versions, and updates PSY status. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The benchmark uses only five very short anime clips and no completed visual comparisons, yet asserts that the testing is perfectly representative of any modern content people typically encode and promotes broad preset recommendations; adding two objective metrics does not support that population-level generalization. | 9/7/2026 | 5 |
| Ogg | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote as a short practical page: extensions table, Opus/Vorbis/Theora commands, lossless Opus remux between MP4 and .opus, Ogg Opus 48 kHz/pre-skip quirks, and Safari 18.4 Ogg support (verified in WebKit's release post). Independent review: Safari 18.4 Ogg Opus/Vorbis support verified on WebKit's release post (March 31, 2025); added the OS versions it applies to. Verified Opus/Vorbis/Theora/FLAC Ogg commands, MP4 Opus -> .opus stream copy, sample_rate=48000 reported for .opus, and that the Ogg muxer rejects MP3. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately covers Ogg pages and logical bitstreams, Opus headers and 48 kHz granule timing, codec-specific mappings, metadata, and appropriate tools. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added pages/packets, logical-stream/chaining and codec-specific granule timing, CRC integrity boundaries, precision ownership, tested Opus encode and WebM-to-Ogg remux, and dated Safari 18.4 Ogg Opus/Vorbis support with exact Apple OS scope. Does not generalize that support to Theora/FLAC/chaining/encoding. | 10/8/2026 | 4 |
| Opus | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrote the opusenc 0.2, FFmpeg libopus and native opus reference rows. libopus -frame_duration: the bound row was correct — libopusenc.c at n9.0.2 accepts exactly 2.5/5/10/20/40/60/80/100/120 ms (80–120 with libopus ≥1.2; the help's continuous range is misleading); confirmed locally (80/100/120 accepted, 30 rejected). Fixed --version-short (same as --version), --discard-comments also drops pictures, --picture accepts GIF, --raw-rate range and resampling, worked --bitrate defaults; native opus needs -strict -2 and is CELT-only. Coordinator: removed code formatting from one option-table header row (contrast fix). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The capabilities, coding modes, channel mappings, rate-control advice, opusenc/FFmpeg distinction, and validation guidance are accurate and appropriately avoid universal bitrate claims. Earlier pass: 2026-09-07GPT-6 Astra · medium Expanded practical guide with version-scoped recipes, controls, tradeoffs, compatibility, and validation. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Adds complete opusenc, libopus and native FFmpeg Opus inventories, defaults, hidden aliases and allowed integer CTL controls. Music, speech/FEC, short-frame CBR, multichannel, raw-pipe and metadata examples passed strict encode/decode checks; transport/FEC recovery is not claimed. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Extends RFC-backed precision, rates and multistream/channel-mapping distinctions while retaining the complete prior parameter inventories. Corrects dated Safari Ogg playback guidance with WebKit 18.4 OS requirements and separates HTML, WebRTC, WebCodecs and recording APIs. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Page was a ~4,800-word reference with repeated 'no universal bitrate' hedging, a 'reproduce the command checks' section and a validation checklist, while omitting the Xiph bitrate recommendations. Rewritten around why to use Opus, Xiph recommended bitrates per channel layout, minimal opusenc/FFmpeg commands, 48 kHz resampling and the 5.1(side)/mapping_family surround pitfall; parameter tables moved under 'Full option reference'. Independent review: Recommended-bitrate table matches the Xiph wiki (VoIP 10-24, podcasts 24 mono/32 stereo, streaming 64-96, storage 96-128, 5.1 128-256, 7.1 256-450). Stereo-to-mono fold-down threshold corrected to ~19 kbit/s per Xiph; added Xiph's FLAC-for-archiving point. 5.1(side) tested in FFmpeg 9.0.2: plain and -mapping_family 1 both fail with 'Invalid channel layout 5.1(side)'; -af aformat=channel_layouts=5.1 fixes both, and SL/SR land in BL/BR (no downmix). Page now says that, explains why -mapping_family 1 is still worth adding (FFmpeg uses the surround encoder API only when asked; default writes a family-1 header via the legacy API), and tells readers to use -filter:a:N when several audio tracks are encoded, matching /guides/audio-and-subtitles. 'Decoder always outputs 48 kHz' corrected (libopus can decode at 8-48 kHz; players and FFmpeg output 48 kHz). '--speech/--music only matter below 40 kbit/s' was unsourced; replaced with opusenc's own description (override detection, low-bitrate effect). Verified: libopus rejects -q:a ('Quality-based encoding not supported'); native 'opus' requires -strict -2; frame sizes 2.5-60 ms with 20 ms default per Xiph. Removed a filler ffprobe/pre-skip line and trimmed repetitive prose around the bound opusenc/--set-ctl-int tables (tables untouched). J1's open issue about -frame_duration 80/100/120: libopusenc.c accepts them when libopus has OPUS_FRAMESIZE_120_MS, so the bound row is right. | 10/8/2026 | 11 |
| PNG | current | PassReview basis· claude-opus-5-5 · high · Correction check oxipng 10.2.1, OptiPNG 7.9.1, pngquant 3.0.3 and ZopfliPNG df1517d verified current. Fixed --keepchunks (default is an empty keep-list, exact comma-separated four-letter names, example corrected), pngquant --no-force and --quiet defaults ('off'; --quiet has no effect), and rewrote all descriptions without hedging. Independent review: No wrong rows in about 35 sampled. oxipng presets, brute lines/levels, threads (Rayon 0 = default) and interlace all match options.rs/main.rs. The pngquant 3.0.3 CLI really is rust/bin.rs (getopts): --no-force defaults off and beats --force, --quiet has no effect, --iebug is ignored, single -Q N gives 0.9N–N, and speeds 10/11 behave as described. OptiPNG abbreviation lists match optipng.c prefix rules, and -log is rejected. ZopfliPNG -m multiplies iterations by 4. The changed defaults (--keepchunks, pngquant --quiet/--no-force) are correct. Coordinator: the appendix now links to the consolidated 'Reading option tables' page. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The help-wanted checklist omits valid 1-, 2-, and 4-bit PNG modes and calls progressive decoding merely 'kinda' even though Adam7 interlacing and progressive display are explicitly standardized; it lacks practical tooling and sources. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Reviews pinned oxipng 10.2.1, OptiPNG 7.9.1, Rust pngquant 3.0.3 and September 30 ZopfliPNG snapshot contracts, hidden aliases and omission defaults. Examples fully decode and distinguish 8/16-bit sample preservation from alpha cleanup, depth scaling, quantization and metadata/APNG changes; separate destinations and refusal/dry-run behavior are checked. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds full-sample color/alpha, metadata, integer-depth and optimizer role distinctions while retaining the complete four-tool parameter references. Dates PNG/APNG support and separates broad decode from third-edition HDR, precision and animation-preservation claims. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): About 4,800 words of option inventories, test narration and an 'Editorial review' caution replaced by a practical guide: PNG properties and the third edition (W3C Recommendation, 24 June 2025), when to use it, oxipng options (-o, --strip safe, --alpha, --scale16), ECT, pngquant lossy palettes with exit codes, and pitfalls (metadata/colour chunks, interlacing, 16-bit, APNG). The four bound tables (oxipng, OptiPNG, pngquant, ZopfliPNG) are kept verbatim as an appendix. Independent review: Confirmed from oxipng v10.2.1 source that --strip safe keeps only DISPLAY_CHUNKS (cICP, iCCP, sRGB, pHYs, acTL, fcTL, fdAT), so gAMA/cHRM are dropped as the page says; -o default 2, max = 6; -z is Zopfli (-Z hidden alias); -a alters hidden RGB; --strip all flattens APNG. Interlacing pitfall suggested `oxipng -i off` as if needed; oxipng's --interlace default is already off, so the pitfall now says oxipng removes Adam7 by default and how to keep it. Verified ECT 0.9.5 (default level 3, -strip, -progressive for JPEG) and pngquant 3.0.3 README (60–80 % figure, exit 99 for quality too low, --speed default 4, --ext .png --force overwrites). Renamed 'photo.png' examples to 'image.png' on a page that says PNG is not for photos. Coordinator: linked the option appendix to the image encoder parameter reference. | 10/8/2026 | 8 |
| Preserve color and HDR | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Expanded a vague stub into the practical colour how-to: identify, preserve, retag (setparams, h264/hevc/av1_metadata bitstream filters, mkvpropedit), convert (zscale/scale/VapourSynth resize for SD BT.601->HD BT.709; full->limited range), a flag table for x264/x265/SvtAv1EncApp/aomenc, a tested 10-bit HDR10 Y4M pipe into x265, HDR10 static-metadata syntax for x265/x264/SVT-AV1, and an HDR10+/Dolby Vision passthrough matrix. Corrected the old advice to tag with '-color_primaries … -color_trc …': in FFmpeg 9.0.2 those output options are ignored and -colorspace/-color_range trigger an auto-inserted conversion. Independent review: Wrong: 'remux with mkvmerge (which reads the new values from the bitstream)' after a bitstream-filter retag. mkvmerge 102.0 keeps the stale Colour elements of a Matroska input (and writes none from a raw stream). Also the stale values affect MP4 colr too, not only Matroska. Replaced with a single tested command: bsf plus -color_primaries/-color_trc/-colorspace output options, which with -c copy set only the container fields; mkvpropedit kept for fixing existing files. The scale-filter SD->HD alternative converted only the matrix but then tagged primaries/transfer bt709, contradicting the page's own 'don't tag without converting' rule. Now keeps the source primaries/transfer tags (smpte170m) and gives a tested command. Added that zscale on yuvj420p input converts the samples to 16-235 but leaves the output tagged pc (tested), so format=yuv420p is needed for both scale and zscale. Caution box scoped to re-encoding (output options do work for stream copy). Added the -x265-params/-svtav1-params form of the HDR10 options (tested). Removed a 'don't prove' phrasing. mkvpropedit property names switched to the canonical color-* spelling. Confirmed by test (FFmpeg 9.0.2): -color_primaries/-color_trc ignored when encoding; -colorspace bt709 converts samples (untagged treated as BT.601) and -color_range pc converts, -color_range tv on yuvj input does nothing; setparams tags only; plain transcode keeps tags, chroma location and HDR10 MDCV/CLL with libx264, libx265, libsvtav1 and libaom-av1 (Changelog: 7.0 for the first three); -x265-params colorprim leaves Matroska without Colour elements; -strict -1 still needed for 10-bit Y4M; XCOLORRANGE=FULL honoured by x264 and aomenc, ignored by x265 4.3 and SvtAv1EncApp 4.2 (fixed with --range full / --color-range full); x265 encodes 10-bit input as Main 8-bit without --output-depth 10. Confirmed HDR10 syntax by encoding and reading back side data: x265 --master-display/--max-cll (x,y x50000, L x10000), x264 --mastering-display/--cll same integer format, SvtAv1EncApp --mastering-display with float x,y and nits plus --content-light, and the same via -x265-params/-svtav1-params. Confirmed dynamic metadata matrix: libx265 and libsvtav1 drop HDR10+, libaom-av1 passes it (Changelog 8.0), -x265-params dhdr10-info restores it; FFmpeg's -dolbyvision defaults to auto but auto did not pass Dolby Vision through in tests, -dolbyvision 1 does (x265 needs VBV; libsvtav1/libaom-av1 write profile 10); MP4 needs -strict unofficial for the dvcC/dvvC box (with stream copy as well).
Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately separates pixel conversion from metadata assignment, cautions against inferred signaling, covers the key color fields and static versus dynamic HDR requirements, and requires both metadata and color-managed visual validation. | 10/8/2026 | 8 |
| Privacy | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Coordinator: forge card metadata now comes from a committed snapshot refreshed by a maintainer script rather than a live build-time fetch; updated the privacy notice and its link to the renamed section. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Updated the native wiki notice for the separate Cloudflare Access-protected Ask preview. Checked deployed code for provider context, browser-tab memory, hashed rate-limit counters and application logging, and described other services’ processing without inventing retention promises. This native page has no migrated-content maintenance block. Earlier pass: 2026-10-05T02:43:58.154Zgpt-6.1-sol · xhigh Updated public Ask hostname and anonymous access description for the owner-authorized production cutover. Checked the current chat, rate-limit, provider and configuration code at 5860b08 and the prepared server policy: OpenCode Zen GLM 5.3 Flash, memory-only conversations, keyed network-address counters and aggregate token/cost persistence. The separate development preview remains protected by Cloudflare Access; provider retention remains subject to its published terms. | 10/8/2026 | 4 |
| ProRes | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced two hedged examples and the RAW/alpha caveat pile with a profile table (FFmpeg profile names, Apple's 1080p29.97 rates) and two minimal prores_ks commands. Fixed the original's data-rate error (it quoted MB/s) and its claim that hardware acceleration started with the M2; Apple lists ProRes engines from the M1 Pro and M1 Max. Independent review: Real pitfall found and documented: prores_ks -profile:v hq with a 4:4:4 source and no -pix_fmt writes a non-standard 4:4:4 stream tagged HQ/apch (verified with ffprobe). The page now explains why -pix_fmt yuv422p10le is needed. Checked all six profile names against FourCCs (apco/apcs/apcn/apch/ap4h/ap4x) and the auto profile (HQ for 4:2:2, 4444 for 4:4:4). Alpha is auto-selected from RGBA input. Explained that ffprobe reports 4444 output as yuv444p12le/yuva444p12le even though prores_ks takes 10-bit input. -vendor apl0: verified it only changes the 4-byte creator ID in each frame header (the MOV sample-description vendor stays FFMP). Added one factual sentence and cut an unverifiable 'QC tools check it' clause. Listed FFmpeg's other ProRes encoders: prores_aw (about 3x faster in a quick test), prores_videotoolbox, and the Vulkan prores_ks added in FFmpeg 8.1 per the Changelog. The ProRes RAW decoder is dated to FFmpeg 8.0. Profile rates match Apple's support page (45/102/147/220/330/500 Mb/s at 1080p29.97). Coordinator: removed inline code formatting from a table header row, which failed WCAG colour contrast on the themed table headers. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked Apple conventional profile precision, alpha and RAW distinctions; corrected unit confusion and M1 Pro/Max hardware history using Apple specifications. Checked prores_ks profile 3 locally and documented independent encoder precision/container limits with a practical encode/decode trial. Locally smoke-tested FFmpeg 9.0.2 HQ output and successful 10-bit 4:2:2 decode. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Repaired truncated description; separated RAW/RAW HQ sensor capture from conventional profile table and scoped the Apple12-bit comparison chart, corrected up-to-16-bit alpha, removed unverified universal editor support, added and tested prores_ks profile 4 alpha command with partial/full opacity, documented 10-bit input versus 16-bit coded alpha and lossy color, and quoted optional audio mapping. | 10/8/2026 | 6 |
| PSNR | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced hedged prose, a '-h filter=psnr' preamble and a settb/setpts/shortest/repeatlast command with a plain statement of what PSNR is for (codec dev, regression, pipeline checks) and why it misjudges psy-tuned encodes (x264/x265 --tune psnr, SVT-AV1 --tune 1, x264's own 'results will be invalid' warning, CID22 correlation). Minimal FFmpeg command, scaling, one-line timestamp note, frame-count pitfall, and how FFmpeg pools the summary (mean MSE then dB, plane weighting); removed testing narrative and reproducibility checklist. Independent review: Lead paragraph was only a definition; added one sentence so a newcomer immediately learns PSNR rewards blur and can't say which encode looks better, and trimmed the following section to avoid repeating it. Added that SVT-AV1's default tune is 1 (PSNR), verified in SvtAv1EncApp 4.2.0 --help and Parameters.md. Verified: CID22 SRCC PSNR-Y 0.62 vs SSIMULACRA2 0.88 (README tables); x264 '--psnr used with psy on: results will be invalid!' (local x264); x265 psnr/ssim tunes disable psy (presets.rst); FFmpeg summary = mean MSE then dB with pixel-count plane weights (vf_psnr.c); libvmaf PSNR caps 60/72 dB (FAQ, 6N+12). Tested: distorted-first psnr with stats_file; mismatched size errors ('Width and height of input videos must be same'); bicubic scale fix; 8-bit reference auto-converted to 10-bit against a 10-bit encode. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added the PSNR definition, format and alignment requirements, and an exact FFmpeg statistics workflow. Corrected the universal encoder-RDO claim and distinguished pooled MSE from averaging frame dB. Synthetic identical and noisy 12-frame inputs produced Inf and 35.792682 dB respectively. | 10/8/2026 | 6 |
| Psychovisual encoding | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Kept the original's x265 and Cloudinary quotes (re-verified against current docs) but fixed errors: x264 has psy-rd/psy-trellis, not psy-rdoq; SVT-AV1-PSY is archived and its AC bias / variance boost work is now in mainline SVT-AV1 (tune 0, --enable-variance-boost, --ac-bias). Added adaptive quantization, x265 psy defaults (psy-rd 2.0, psy-rdoq 0/1.0 at slow presets), why x264 --tune psnr/ssim disable psy, and the community stance that metrics penalize psy tuning so viewing decides. Rewrote the garbled fidelity-vs-appeal paragraph. Independent review: Verified x264 psy-rd 1.0:0.0 and aq-mode defaults, x265 psy-rd 2.0 and psy-rdoq 0 with 1.0 at slow/slower/veryslow/placebo (param.cpp), SVT-AV1 v4.2.0 tune 0=VQ/1=PSNR default, --enable-variance-boost and --ac-bias (Parameters.md); consistent with /tools/software-encoders/svt-av1/. Confirmed SVT-AV1-PSY repository archived (Feb 2026); linked its wiki page. Metric paragraph matches the brief's consensus and links /guides/compare-encoding-quality/. | 10/8/2026 | 4 |
| qaac | current | PassReview basis· claude-opus-5-5 · high · Correction check Checked against options.cpp/options.h/main.cpp at v3.07: default TVBR 90 (CVBR for --he), --quality 2. Fixed --no-optimize (parsed but unused in 3.x), --artwork-size also resizes copied art, --stat output, --threading/--exec behaviour; --tvbr row now lists Apple's 15 steps (90 encodes as q91). All cells rewritten. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Coordinator: TVBR values snap to 15 steps, so 90 and 91 encode identically; wording no longer presents 91 as higher quality. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks qaac 3.07 registered long options, short-only controls and custom exec parser against versioned source. Separates frontend defaults from proprietary Apple capability choices and labels all recipe validation source-only; no runtime encode or performance result is claimed. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Page was only a 76-option reference plus 'not executed' disclaimers. Added what the tool is, Windows-only status, how to obtain Apple Application Support, TVBR values with typical bitrates, the CoreAudioToolbox CBR glitch notice from the 3.07 README and common pitfalls; option tables kept verbatim as the reference. Latest release confirmed as 3.07 (2026-09-10). Independent review: qaac 3.07 (2026-09-10) confirmed as latest. TVBR table matches the cited Doom9 thread (127~320, 109~255, 91~195, 82~165, 73~150, 64~135; 5.1 at V91 ~450). Default -V90, --quality 0-2 default 2, --he forbids TVBR, stdout only for ADTS/WAV all confirmed in options.cpp. CBR glitch notice reworded to match the README exactly (7.9.8.x up to 7.10.9.0, only found in CBR) instead of asserting other modes are unaffected. -r keep is the default (options.h rate(-1)); clarified -r auto lets the encoder choose. Installation steps match the qaac wiki (no Microsoft Store iTunes). Coordinator: removed inline code formatting from a table header row, which failed WCAG colour contrast on the themed table headers. | 10/8/2026 | 6 |
| QOI | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrote descriptions; corrected the claim that extra qoiconv arguments are ignored (a leading flag is read as the input path and fails; only arguments after the second operand are ignored); height limit stated as width×height below 400 million pixels. FFmpeg's qoi encoder has no private options. Independent review: All rows correct: qoiconv argv handling, forcing non-3-channel PNG to RGBA, and the qoi_desc validation (non-zero size, channels 3–4, colorspace <= 1, height < 400M/width). Coordinator: the appendix now links to the consolidated 'Reading option tables' page. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Reviews the pinned reference converter's two positional operands, explicit library encode contract and FFmpeg 9.0.2 zero-private-option scope. Tests compare RGB/RGBA including hidden samples, verify frame-plus-codec linear labeling versus a source-overridden output tag, and avoid claiming PNG metadata, higher precision or original file-byte preservation. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds the specification-bounded 8-bit RGB/RGBA, alpha, header and single-image model, plus reference/library/FFmpeg decode paths. Preserves converter and API references and distinguishes custom JavaScript/Wasm display from assumed native browser support. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Replaced the converter/library contracts and exact-roundtrip narration with a short practical page. Corrected the original import's claim that QOI beats libpng on size: the author's benchmark (2,879 images) shows QOI about 16% larger than libpng but about 30x faster to encode and 3x faster to decode. Restored the chunk-type summary and added usage for qoiconv, FFmpeg and ImageMagick. The three bound tables/sections are kept as an appendix. Independent review: Benchmark table matches qoiformat.org/benchmark 'Grand total' exactly (2,879 images, i7-6700K, single-threaded; libpng 398 KB/83.8 ms/7.0 ms, stbi 561 KB, QOI 463 KB/2.9 ms/2.1 ms), so '16 % larger, ~30x faster encode, ~3x faster decode' is correct. Removed the meta sentence about what 'the original wiki text said' (wiki self-narration). FFmpeg example dropped redundant `-c:v qoi` (the .qoi extension selects the encoder); round trip tested pixel-identical. Coordinator: removed the ignored -color_trc output option; setparams alone sets QOI's linear header flag (verified locally). Coordinator: linked the option appendix to the image encoder parameter reference. | 10/8/2026 | 6 |
| QSV | current | PassReview basis· claude-opus-5-5 · high · Correction check Added descriptions for all 17 av1_qsv options (preset 0 = runtime default and overridden by -compression_level; b_strategy maps to BRefType pyramid/off; forced_idr makes every I-frame a key frame; look_ahead_depth needs -extbrc 1) and tightened the intro. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page gives generation-independent codec claims and universal vendor-quality rankings without evidence, and still centers the superseded Intel Media SDK instead of current oneVPL capability discovery. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced generation-independent codec and vendor-quality assertions with Intel VPL dispatcher/runtime capability discovery, identified Media SDK as legacy, and added a device-qualified h264_qsv baseline. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks all 17 FFmpeg 9.0.2 AV1 private option spellings, ranges/named values and wrapper defaults, plus codec-specific generic initialization overrides. Hardware recipes are independently source/help-checked, not GPU-executed; driver negotiation, automatic settings and compatibility remain explicit limits. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Removed the discovery preamble, the H.264 5M baseline and the 'source-checked, not executed' prose. Added hardware/codec support (AV1 encode on Arc and Meteor Lake+, Tiger/Alder/Raptor Lake decode only), libvpl >= 2.6 and vpl-gpu-rt requirements, ICQ via -global_quality as the quality command, the CQP-by-default pitfall (QP 26, AV1 128), -look_ahead for H.264 only, bitrate mode selection, vpp_qsv full-GPU pipeline and an options table. Marked parameter block moved verbatim to the appendix. Independent review: Hardware table corrected from Intel's media-driver feature table: HEVC 8-bit encode starts at Skylake (page said Kaby Lake); the Linux media driver exposes HEVC 10-bit from Ice Lake; VP9 encode is Ice Lake+ (8/10-bit, low-power only, forced by qsvenc_vp9.c) and absent on Lunar Lake and Battlemage (page claimed Kaby Lake 8-bit). AV1 on Alchemist/MTL/LNL/BMG/PTL confirmed, which resolves batch O's open issue. Rate-control text aligned with select_rc_mode(): ICQ is chosen when -global_quality is set without -maxrate (not 'without a bitrate'); -b:v without equal -maxrate is VBR; -global_quality plus a VBR setup gives QVBR. Default-CQP QP 26 / AV1 128 and -look_ahead being H.264-only verified. libvpl >= 2.6 verified in configure; added that libmfx is deprecated and av1_qsv depends on libvpl. vpl-gpu-rt platform list (Tiger Lake+) verified. -look_ahead_depth max 100 and extbrc applying to CBR/VBR noted from source. | 10/8/2026 | 5 |
| Quality controls: CRF, bitrate and presets | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Rewrote from a short abstract table plus a 'reproducible comparison' checklist into a correct explanation of CRF/constant quality, constant QP, one-pass ABR, two-pass VBR, CBR and VBV-capped CRF, with a per-encoder CRF table (defaults: x264 23, x265 28, SVT-AV1 35, libvpx 32; FFmpeg's libsvtav1 -crf capped at 63), a tested capped-CRF command, a file-size bitrate formula and tested two-pass commands. Corrected the folklore that libvpx-vp9 needs -b:v 0: in current FFmpeg -crf alone selects constant-quality mode. Independent review: CRF table listed libaom's default as '—'; FFmpeg's libaom-av1 falls back to CRF 32 when neither -crf nor -b:v is set (libaomenc.c, confirmed by encoder warning). Fixed. Verified the rest of the table: x264 23, x265 28 (0–51), SVT-AV1 35 with 1–70 range and FFmpeg's 0–63 AVOption cap, libvpx-vp9 32 and -crf alone = constant quality (set_vp9_defaults), libaom -crf + -b:v = constrained quality. VBV was described as 'limits how fast the decoder's buffer may fill'; rewrote as the buffer model (fill at -maxrate, size -bufsize) and noted -maxrate/-bufsize also map to libx265 VBV and turn libsvtav1 CRF into capped CRF (SVT log shows 'capped CRF'). Tested capped-CRF x264 command and both two-pass commands (stats file ffmpeg2pass-0.log); size formula arithmetic (1500 MiB/60 min − 128k ≈ 3367k) checked. SVT-AV1 multi-pass statement matches CommonQuestions.md. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page correctly frames rate control around delivery constraints, warns that CRF and presets are encoder-specific, and accurately describes SVT-AV1 multi-pass as chiefly useful for VBR targets and generally noncritical for ordinary CRF work. | 10/8/2026 | 6 |
| QuickTime File Format (MOV) | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote as a practical MOV page (ProRes/DNxHR/PCM/timecode editing masters, MOV vs MP4 relationship, hvc1 tag, colour tags); removed the generic MOV-vs-MP4 hedging and testing narrative. Independent review: Pitfall said 'ProRes, PCM and timecode tracks do not go into MP4'. Tested: FFmpeg's MP4 muxer rejects ProRes and DNxHD/DNxHR but writes PCM audio and a tmcd timecode track. Rewritten accordingly. ProRes 422 HQ bitrate sentence was garbled ('at 1080p ... at 1080p24–30'); now 'about 176–220 Mb/s at 1080p23.976–29.97' (Apple white-paper figures). Verified prores_ks profile numbers, that profile 4 with 4:2:2 input yields a 4:2:2 file labelled 4444 (so the pix_fmt note is needed), dnxhr_hq to MOV, -timecode tag. Coordinator: added -pix_fmt yuv422p10le to the ProRes 422 HQ example after reviewer RH2 showed prores_ks keeps 4:4:4 input as a non-standard stream labelled HQ; tested with FFmpeg 9.0.2. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The concise page accurately explains QuickTime File Format's relationship to ISOBMFF, its production uses, non-interchangeability with MP4, and metadata-preservation risks. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Expanded editing inspection and selected-track remux guidance, separating codec precision/chroma/alpha from MOV, edit-list timing from damage, and supported global metadata/timecode generation from original private-atom/data-track preservation. ProRes 422 HQ/PCM fixture remux and full decode passed; no editor or timecode-preservation certification claimed. | 10/8/2026 | 4 |
| rAV1ator | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is explicitly incomplete, has no canonical source or reviewed version, and recommends installing from an unsigned Flatpak remote with --no-gpg-verify without provenance or risk guidance. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced stale installation guidance with canonical EOL/archive status, existing-install inspection and migration/reproduction workflow; removed unsigned-remote recommendation. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Cut the Flatpak forensics/reproduction instructions; now a short end-of-life page: what rAV1ator was (Av1an + rav1e GUI forked from Aviator), archive status (EOL, archived October 2024, last release 1.1.2), and what to use instead. Independent review: 'Forked from Aviator' was unverifiable; now described as Aviator's sister project. Verified archive (2024-10-20), last release 1.1.2 (2023-05-16), Trix (nekotrix) as contributor, personal Flatpak repo instead of Flathub, Av1an+rav1e. | 10/8/2026 | 1 |
| rAV1ator CLI | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Distinguished the archived original and the successor’s reviewed July 2025 revision from assumptions about ongoing compatibility. Qualified historical bundled help, WSL guidance and hash checks, repaired the internal article route and removed blanket image-removal claims. Interactive installation and updating were not executed. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): The page showed a stale v0.2.4 help transcript and carried a ~1,500-word WSL screenshot tutorial, a body H1 and an unused Tabs import. Rewritten short: what it is, maintenance status (original archived January 2025; Ultimaxx fork v0.4.0, no commits since July 2025), install (must be in /usr/local/bin), current dependency list from the v0.4.0 script, a note that the Arch av1an release package fails with current VapourSynth, and a verified option table; WSL now links Microsoft's docs. Independent review: Verified handover to ultimaxx (original archived 26 Jan 2025), fork v0.4.0 (2025-02-02), last commit 2025-07-19, and every option in the table against the v0.4.0 script's help. | 10/8/2026 | 4 |
| rav1e | current | PassReview basis· claude-opus-5-5 · high · Correction check Rechecked all 50 native rows against rav1e v0.8.1 (common.rs, rav1e.rs, speedsettings.rs, rate.rs, levels.rs, color.rs, Cargo.toml) and local rav1e 0.8.1, and all 6 librav1e AVOptions against FFmpeg n9.0.2 librav1e.c and ffmpeg -h. Fixes: --rdo-lookahead-frames row copied rav1e's wrong help text (speed 6-8 = 20, 9-10 = 10); the preset code sets 20 at 6-7 and 10 at 8-10. --slots default was given as 'automatic'; 0/1 means off, and only the rav1e-ch binary (unstable + channel-api) uses it. --quantizer default is now '100 (255 with --bitrate)', and the row says 0 (lossless) aborts. --level auto writes seq_level_idx 31, the same as unconstrained (confirmed with trace_headers); the row lists the levels that are accepted (2.0-6.3; 7.x rejected). --frame-rate needs --time-scale (time-scale 0 fails). --range ignores Y4M and defaults to Limited. The advanced -s/-l 'conflict' claim was dropped because clap does not enforce it. save/load-config need the non-default serialize feature, --load-config replaces CLI encoder options, and `advanced -c zsh` panics in 0.8.1. Also: the output must be .ivf, -y matters for piped input, and photon noise is luma-only at ISO = strength x 100. Regrouped 50 rows into 9 real groups, rewrote every description, and rewrote the intro. FFmpeg: the defaults were correct. Rewrote the rows and intro to cover what the source does: -rav1e-params is applied before every wrapper/generic option, -speed (and a speed key) resets rdo_lookahead_frames/no_scene_detection, rav1e-params booleans must be true/false, rav1e-params bitrate is bit/s, bad keys only warn, -tiles overrides rows/columns, and the generic -g/-keyint_min/-b:v/-qmin/-qmax/-pass/-threads mapping is listed. Both snapshots verifiedAt 2026-10-08; native sourceFiles expanded. Prose fix by T2: --quantizer range is 1–255 (0 aborts; verified with rav1e 0.8.1). Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08).
Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately pins v0.8.1, distinguishes CLI/library/FFmpeg interfaces, documents supported input depth and chroma, and avoids unsupported universal settings. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks rav1e 0.8.1 main/advanced CLI inventory, 49 stable rows plus a labeled unstable-feature row, and six FFmpeg 9.0.2 private options. Native/wrapper samples passed strict decode and expected frame counts; interface and build-feature limits stay explicit. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Replaced a padded, checklist-heavy page with a short guide. New 'Where rav1e stands' section: still v0.8.1 (June 2025), weekly pre-releases stopped Sept 2025, only maintenance commits since, slower than SVT-AV1 per cited benchmarks. Restored the original's practical guidance, corrected: input is positional (`-i` is min-keyint, so the original `rav1e -i input.y4m` was wrong), tiles are `--tile-cols`/`--tile-rows` counts (the original's `--tile-columns` doesn't exist), ×4 cq-level→quantizer tip restored with the qindex explanation. Added HDR10 signalling, a pointer to the still-unmerged HDR10+ PR #3000, and the fact that FFmpeg librav1e cannot do photon noise (not a C API key). Removed the per-workflow 'experiment' sections and testing narrative. Option tables kept verbatim in an appendix. Independent review: Two-pass example was broken: the first pass writes the -o file too, so the second pass stops at an 'Overwrite? [y/N]' prompt (exits when non-interactive). First pass now writes to /dev/null, with one sentence explaining why. RDO lookahead values were wrong ('20 at 6–8'); speedsettings.rs at v0.8.1 gives 40 (0–2), 30 (3–5), 20 (6–7), 10 (8–10). Fixed. Removed the libheif 2022 benchmark figures (single still image, luma-PSNR only, rav1e 0.5) as weak evidence for video; kept the mendevi dataset conclusion with a link and dropped the libheif source. Verified: v0.8.1 (2025-06-16) is latest; weekly pre-releases end at p20250902; later commits are lint/MSRV/dependency upkeep; positional input and -i = min-keyint; tile options are counts; --tiles conflicts; C API key list matches capi.rs; FFmpeg -rav1e-params photon_noise is rejected with a warning and the encode continues; NASM 2.14.02 requirement. Tested: file encode, 10-bit pipe (FFmpeg 9.0.2 still needs -strict -1), tiles+photon noise, corrected two-pass, HDR10 signalling (ffprobe shows smpte2084/bt2020), FFmpeg librav1e. | 10/8/2026 | 8 |
| Reading option tables | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Coordinator: merged the three parameter reference pages into one 'Reading option tables' page: where each tool's full option table lives, what the Default column means (checked against FFmpeg options_table.h and the Gifsicle manual), and FFmpeg's generic encoder options. The image and processing-tool parameter pages were retired with redirects. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Provides a discoverable interface/version index and distinguishes fixed, automatic, preset/input-dependent, unset and unknown defaults. Separates encoder-private from generic FFmpeg controls, corrects rate versus buffer units and explicitly labels proprietary and hardware verification gaps. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-haiku-5-5): Removed the 'ffmpeg -version / -h full' preamble block and the 'Build a reproducible example' checklist and its verification narration ('source-checked', 'does not prove', 'a successful…'). Kept the encoder inventory table and the default-label and generic-output-control tables unchanged, and added a one-line example of `ffmpeg -h encoder=libsvtav1` in place of the preamble. Frontmatter sources, lastVerified and license are unchanged; description rewritten. Independent review: Reworded two robotic cells in the generic-options table (-q:v/-crf scales, -threads). Validator passes (3337 rows); bound data untouched. | 10/8/2026 | 3 |
| Reducing Image Load Online | current | PassReview basis· GPT-6 Astra · medium · Correction check Corrects lazy loading, CSS/device-pixel sizing, universal codec/XYB compatibility claims and JPEG reconstruction; keeps image files, historical commands, scores/sizes/timings and attribution while limiting inference to one photograph. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page incorrectly says native image lazy loading does not save bandwidth, even though deferred images may avoid network transfer entirely when never approached; it also mixes one-image codec results and opinionated WebP/JPEG claims into broad format recommendations without sufficient evidence. | 9/7/2026 | 6 |
| Reference | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Coordinator: links now point to the consolidated 'Reading option tables' page instead of the retired parameter pages. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-haiku-5-5): Rewrote the section intro into a direct guide to the parameter pages and the browser/format-support pages, and tightened the description. Link targets were checked to resolve; the summaries of what each parameter page covers were taken from those pages' tables and not re-verified here. Independent review: Checked every link target exists and that the encoder-parameters summary matches that page's table; removed a generic closing 'play a real sample before you ship' sentence. | Not recorded | 0 |
| Resources | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Coordinator: links now point to the consolidated 'Reading option tables' page instead of the retired parameter pages. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Separated first-party specifications and tool manuals from community filtering guides and historical recipes. Identified incomplete guide coverage and unstable APIs, and avoided treating older recipes as verified current CLI documentation. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-haiku-5-5): Removed the 'inspected on 4 October' and 'not freshly validated' narration and the 'explicitly warns' filler; the lvsfunc warning is now stated with its actual wording. The H.273 row now names the in-force edition (H.273 07/24) verified on the ITU page, and the Xiph row names the projects the page lists. lastVerified updated to 2026-10-08 after all links were fetched. Independent review: Verified H.273 07/24 is the newest edition listed by ITU, the lvsfunc 'Anything MAY change at any time' warning and the Fansubbing Guide 'under construction' note. Kokomins entry now says what it is useful for (psy-rd, psy-rdoq, AQ, deblock, no-sao, all present in the article) instead of a generic 'check against your source' hedge; x264 settings link (HTTP 200) kept with a concrete pointer to x264 --fullhelp. | 10/8/2026 | 9 |
| Spatial and immersive audio | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Removed browser-testing prose and the workflow checklist; kept the layered explanation (codec vs format vs renderer vs binaural), a compact comparison table, and the IAMF section (anchor retained). Updated IAMF to v2.0 and codec list from the spec. Independent review: IAMF v2.0.0 (Final Deliverable, 24 Aug 2026) and codec list Opus/AAC-LC/FLAC/LPCM, ambisonics and objects, ISOBMFF and libiamf confirmed in the spec. DTS:X row 'carried by DTS codecs' made concrete: DTS-HD MA extension (FFmpeg profile 'DTS-HD MA + DTS:X'). Auro-3D row 'Auro's own codecs' was wrong: Auro height channels are folded into a 5.1/7.1 PCM signal, usually wrapped in DTS-HD MA on Blu-ray. Replaced vague DTS:X/Auro prose with the practical point (backward-compatible mix; FFmpeg decodes only that; copy to keep). Removed 'test with a known fixture on the real device' and 'test your target application' generic advice; 'channel bed' wording aligned with the Dolby page. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately distinguishes codecs, immersive models, branded ecosystems, and renderers; its IAMF codec list and its qualified comparisons are supported by current first-party specifications. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds scene-aware browser and representation validation, distinguishing base codec decoding from IAMF/Atmos rendering, panning, object metadata and output layout. Identifies channel/object fixture and mix-presentation checks without claiming they were executed. | 10/8/2026 | 5 |
| Speex | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Reduced to: obsolete, replaced by Opus (per speex.org), modes by sample rate, speexenc/speexdec and FFmpeg libspeex commands with the wrapper options; dropped delay/profile prose. Independent review: Verified speex.org obsoleted-by-Opus statement, Speex 1.2.1 as last release, speexenc --wideband --quality 8 / speexdec, and FFmpeg libspeex -b:a 24k; FFmpeg wrapper complexity default is 3 (debug log), frames_per_packet 1-8, vad/dtx/abr exist. No edits needed. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This is only a help-wanted stub and replacement claim; it omits Speex modes, rates, framing/Ogg mapping, tools, migration limits, and sources. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checks all five FFmpeg n9.0.2 private wrapper controls and advertised rates, formats and channels against source and installed help. Adds mode/delay and legacy browser boundaries; the wideband example fully decodes, while unexecuted combinations remain labeled. | 10/8/2026 | 4 |
| SRT | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Cut the repetitive disclaimers; kept format syntax, the comma-vs-period timestamp difference, tag support, charset fixing (-sub_charenc/iconv), muxing commands and conversion losses. Independent review: '{\an8} is honoured by mpv and libass-based players; other players show it as text' was wrong (VLC and most desktop players honour it); reworded. Removed 'lossless in practice' tic. -itsoffset, -sub_charenc CP1252, iconv, file -i, mkv/mp4 mux and extraction verified. Earlier pass: 2026-09-08GPT-6 Astra · not recorded Adds complete UTF-8 authoring, parser/styling limitations, known-charset conversion, Matroska muxing and timing diagnosis; exercises accented Windows-1252 text conversion and a 500 ms cue shift while separating syntax checks from caption quality and accessibility. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked timed Unicode text, renderer-specific styling, Matroska mapping and native HTML WebVTT delivery distinction. Existing synthetic execution evidence remains dated separately; accessibility/editorial review remains pending. | 10/8/2026 | 3 |
| SSIM | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Removed check-build preamble, padded filter chain, disclaimers and testing narrative. Now states up front that SSIM beats PSNR but still rewards smoothing and is mostly luma, with CID22 correlations and encoder tune behaviour, gives the minimal FFmpeg command, explains the dB value and pixel-count plane weighting, notes implementation differences (libvmaf float_ssim downsampling), and adds MS-SSIM via libvmaf's float_ms_ssim feature. Independent review: Added the 'don't judge encodes by it' point to the opening paragraph; tightened the 'good for' paragraph. Corrected 'ssim.log has the same per frame': per-frame lines give Y/U/V/All with dB only for All. Verified: CID22 SRCC SSIM 0.76 and MS-SSIM 0.76 vs SSIMULACRA2 0.88; All = pixel-count weighted (reproduced 0.6139 from Y/U/V); dB = -10*log10(1-SSIM); libvmaf float_ssim downsampling difference (FAQ); x264 '--ssim used with psy on' warning. Tested: ffmpeg ssim with stats_file; libvmaf feature=name=float_ms_ssim JSON log (pooled float_ms_ssim present). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Explained implementation and scale dependence and distinguished SSIM from FFmpeg’s reported SSIM dB. Added an aligned FFmpeg statistics workflow. Synthetic identical and noisy 12-frame inputs produced 1.000000 and 0.922726 respectively. | 10/8/2026 | 8 |
| SSIMULACRA2 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Was the raw migrated page (truncated description, duplicate H1, unused Tabs import, long step-by-step algorithm, a score table that did not match upstream). Restructured: what it is and why the community uses it, the exact upstream score scale (90 visually lossless/flicker test, 85, 80, 70, 50, 30, 10, negative), image scoring with libjxl's ssimulacra2 (ships in libjxl release archives, not most distro packages) and fssimu2, video scoring with FFVship/Vship and vszip (correct property names), summarising with mean/5th percentile/worst frames, and a five-bullet 'how it works'. Dropped archived ssimulacra2_rs. Independent review: Score table paraphrased the README (70 row omitted 'barely noticeable side by side'; no 100 row). Rebuilt it from the upstream README wording and added its reference points (cjxl -d / libjpeg-turbo quality per level). Added that FFVship guesses missing colour tags and that --verbose reports them: the same untagged 640x360 pair scored -25.7 in FFVship (which assumed BT.470BG matrix/transfer/primaries) and -14.0 through the Vship VapourSynth plugin, while correctly tagged BT.709 files gave identical results in both (-14.37). Verified 2.1 (April 2023), XYB/6 scales/54 maps/108 norms, tuning datasets, and that SSIMULACRA2 beats PSNR, SSIM and VMAF on all four README datasets; confirmed Arch's libjxl package ships no ssimulacra2/butteraugli_main while the v0.12.0 static archives do; fssimu2 conformance (mean diff -0.51) supports 'scores differ slightly'. Tested: FFVship default SSIMULACRA2 with --json; the page's VapourSynth script (with the plugin loaded explicitly for the test). | 10/8/2026 | 6 |
| Stabilizing | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Restored the original's practical two-pass vid.stab usage and examples, dropped the outdated accuracy-speed claims, and added the documented unsharp recommendation, a verified option table, limits (rolling shutter, parallax) and alternatives (deshake, Gyroflow). Independent review: First command used smoothing=30 without reason; now the documented default chain (vidstabtransform,unsharp=5:5:0.8:3:3:0.4), with smoothing=30 kept for the very-shaky example. Stated smoothing default 15 from the FFmpeg 9.0.2 AVOption table (the HTML docs still say 10); both two-pass command pairs run. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page mischaracterizes vidstabdetect shakiness as movement-reduction strength and gives unexplained 21/14 fps performance figures without hardware, input, build, or command provenance. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected VidStab shakiness to expected camera-motion magnitude rather than stabilization strength; removed unproven 21/14 fps claims and added matched detection/transformation passes with smoothing and zoom tradeoffs. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. | 10/8/2026 | 3 |
| Start here | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the 'will be migrated with attribution and review status' placeholder with a one-paragraph introduction and an ordered reading list of the seven start-here pages, leading into the guides. Independent review: Reading list matches the seven pages; links resolve. | 10/8/2026 | 0 |
| SubStation Alpha and ASS | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote the ASS page around file structure, colour order (&HAABBGGRR), the tags people use, attaching fonts (verified FFmpeg's subtitles filter loads attached fonts), burn-in and conversions; removed boilerplate about renderer behaviour and the review-status admonition. Independent review: YCbCr Matrix advice ('use TV.709 for HD') was incomplete; libass guide says match the video's range/matrix for BT.601/709 and use None otherwise (e.g. HDR). Fixed. ffmpeg -dump_attachment works but exits with 'At least one output file must be specified'; noted that and added the mkvextract alternative. Colour order, ScaledBorderAndShadow, override colour form checked against the libass guide; attach commands (ffmpeg, mkvmerge, mkvpropedit) and subtitles=video.mkv font loading verified. Earlier pass: 2026-09-08GPT-6 Astra · not recorded Adds complete ASS script geometry, timestamp/color conventions, font inventory, Matroska attachments and libass SDR burn-in; exercises the script at 1280×720 and verifies a switchable ASS/font track and changed rendered pixels, without certifying complex typesetting or HDR. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked SSA/ASS script types, rich styling, libass/Aegisub authoring, Matroska fonts and browser derivative limits. Chroma/precision belong to composited video rather than subtitle syntax; HDR render verification is not claimed. | 10/8/2026 | 5 |
| Subtitle formats overview | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced the 'standards vs implementation' prose and preservation checklist with a format-chooser table, a container-support table, tested conversion/mux/burn-in commands, and the real losses (ASS->SRT leaves <font> wrappers and {\an8}; ASS->WebVTT drops positioning). Documented that force_style FontSize is in 288-line script units. Independent review: Bitmap burn-in was untested. Built a PGS .sup by hand, muxed it and tested: '[0:v][0:s:0]overlay' extends the output to ~4,294,968 s (the last PGS event has no end time) and an MP4 mux fails with 'pts/dts pair unsupported'. 'overlay=shortest=1' gives the exact frame count and correct on/off rendering (also verified with dvdsub converted from PGS). Replaced the one-liner with a tested command block and explanation; added scale-after-overlay note and overlay filter source. ASS->SRT/WebVTT losses verified on a sample: SRT gets <font size color> wrappers, \pos dropped, {\an8} left as text; WebVTT keeps italics and bold and drops the rest. Wording made exact. force_style FontSize: FFmpeg's default ASS header for SRT uses PlayResY 288, so FontSize 24 = 1/12 picture height; claim kept. subtitles= with MKV loads attached fonts ('Loading attached font' in debug log). -sub_charenc and charenc= tested; mkvmerge/ffmpeg mux commands tested. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately separates standardized timed text, de facto text formats, bitmap subtitles, broadcast captions, container mappings, conversion loss, and accessibility requirements. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked subtitle overview links and distinction among simple text, authored ASS appearance and external browser WebVTT tracks. Preserved pending review and original format/container evidence. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted optional maps, added -n, and made ASS burn-in's libass/libx264/video settings explicit with a Matroska derivative, copied-audio/player boundaries and trial-quality wording. Synthetic extraction, mov_text mux and ASS burn-in/decode passed; MKVToolNix examples explicitly remain documentation shapes. | 10/8/2026 | 6 |
| SVT-AV1 | current | PassReview basis· claude-opus-5-5 · high · Correction check Native table (v4.2.0, 143 options) regenerated from app_config.c/enc_settings.c/enc_handle.c and regrouped into 11 purpose groups with every description rewritten. Corrections: --tbr default is 2000 kbps (the old row showed the internal 2000513 bit/s value; --help's 7000 is stale) and accepts b/k/m suffixes; --min-qp/--max-qp are 0–63, and min-qp auto resolves to 0 for CRF and 4 for VBR/CBR; --lp is a 1–6 level with 0 = auto (was shown as '0, 6'); --fast-decode no longer refers to 'the fork'; q-index offsets are −64–63 (validation, not help's −256–255); --scd is switched back off with a warning; presets above 11 run as 11 outside RTC; an .obu output extension selects OBU output; CBR needs low delay and VBR random access; recode-loop auto levels; undershoot auto 50 VBR/25 CBR; --force-key-frames default is none; --tier and --intra-period documented as parser-only. FFmpeg libsvtav1 section rewritten from n9.0.2 source: unset sentinels, -qp disables AQ, -b:v/-maxrate/-qmin/-qmax/-bufsize/-g/+cgop mapping, -svtav1-params applied last and only warns on bad keys. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The v4.2.0 release, standalone input constraints, interfaces, build, rate-control controls, and validation guidance are consistent with the tagged AOMedia documentation. Earlier pass: 2026-09-07GPT-6 Astra · medium Expanded the already-passing mainline page with CRF/preset comparisons, native Y4M piping, two-pass VBR, grain controls and GOP guidance; explicitly distinguished mainline 4.2 tune 5 VMAF from HDR tune 5 Film Grain and local 4.1 tests. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks 143 SVT-AV1 4.2.0 native option rows and five FFmpeg 9.0.2 private options. Resolves stale documentation defaults using initialization code, separates sentinels and mode-dependent values, and verifies mainline quality, two-pass, low-delay and grain examples by strict decoding. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Replaced ~230 lines of repetitive prose (four overlapping guidance sections, 'What was exercised locally', page-licence admonition, reproducibility checklist, --lp 2 / -map / -an / pipefail padding, build-check preambles) with one reference: what it is, getting it (FFmpeg vs SvtAv1EncApp, tagged build), tested starting commands, one options table, a verified tune table, grain synthesis, bitrate mode and pitfalls. Restored community guidance from the original import (preset ranges, film-grain 8/4, 10-bit output, no scene-cut keyframes/Av1an) and added mainline's PSY-derived options. The v4.2.0 release date (2026-07-14) and the tune 0–5 table were verified as correct. New findings were added: tune 5 applies unsharp-mask preprocessing, tunes 2/4/5 are CRF-only, tune 3 is all-intra only, presets 12–13 are remapped to 11 outside RTC, FFmpeg rounds -crf 27.5 to 28, and CBR requires low-delay. Bound FFmpeg/native parameter tables kept verbatim as the appendix; only their surrounding prose was shortened. Independent review: Build command was wrong about `native`: Build/linux/build.sh calls `parse_options enable-native` before parsing arguments, so -march=native is the default; removed the redundant `native` and told readers to add `disable-native` for a portable binary (was 'drop native for a portable binary'). LTO default qualified (GCC 9+/Clang 12+, CMakeLists.txt). Overstated threading claim ('a single encode keeps a modern CPU busy') replaced with upstream's figure: about 16 cores at 1080p presets 4-6, chunking beyond that (CommonQuestions.md). keyint default clarified: 5 s rounded up to whole mini-GOPs (compute_default_intra_period -> 160 frames at 24 fps). --tf-strength row claimed lower values risk 'motion artefacts' (unsourced); rewritten to what upstream says (strong filtering smooths texture/grain). --enable-qm/--qm-min row rewritten to upstream semantics (lower level = steeper matrix, smaller file, slightly lower quality at same CRF, more noticeable at low CRF). Added --variance-octile row (default 5, upstream recommends 4-7, PSY used 6) since it is a standard community knob. --fast-decode pitfall now notes upstream's warning that it was tuned with tune 1. Verified unchanged: SCD forced off with warning in enc_settings.c (claim correct); tune table 0-5 incl. tune 4 = 'MS_SSIM and SSIMULACRA2 optimized' with overrides (enc_handle.c) and tunes 2-5 CRF-only; tune 0 forced to 1 in low delay; variance-boost strength guidance (Appendix-Variance-Boost.md); --ac-bias 1.0-1.5 / 4-6 with TF+CDEF off (Parameters.md); --qp-scale-compress-strength levels; film-grain 8 / 10-15 / ~4 is upstream CommonQuestions.md text; presets 12-13 mapped to 11 outside RTC (enc_handle.c, and observed 'Preset M12 is mapped to M11'); FFmpeg -crf 27.5 -> 28 (observed); keyint=10s via -svtav1-params works with FFmpeg 9.0.2; CBR requires low delay; v4.2.0 dated 2026-07-14 in CHANGELOG. Starting command (preset 4, CRF 30, 10-bit, tune 0, variance boost, keyint 10s) re-run and now identical to the guide; it matches upstream's crf=30 1080p starting point and tune 0 advice. | 10/8/2026 | 8 |
| SVT-AV1-Essential | current | PassReview basis· claude-opus-5-5 · high · Correction check v4.0.1-Essential is still the latest tag. --speed and --preset defaults now depend on resolution (slow/preset 4 up to 1080p, medium/5 above); --qp/--crf default 30 up to 1080p and 35 above when --quality is unset; --quality mapping given in the row; --input-depth: 8-bit Y4M/YUV is rejected (verified with a test encode); --output defaults to the input name plus _essential (verified); --lookahead is 0–32; preset remapping above 11; recode-loop, tune 0–4 and Essential-only options described from source and README. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh Release and branch identifiers, integrated FFMS2/WebM behavior, key feature list, changed defaults, and strong cross-fork caveats agree with the official v4.0.1-Essential material. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks the complete fork-specific registered inventory (157 rows at 4.0.1-Essential revision 7b064486c2f3), default initialization and authored parameter explanations against pinned source. Adds practical native/pass-through combinations without borrowing mainline defaults; fork examples remain source-checked rather than newly binary-tested. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Page listed 'project claims' without values and suggested FFMS2/WebM are part of normal use; the official release binaries are built without FFMS2, WebM, Dolby Vision or HDR10+ (release CI workflow), so they take Y4M and write IVF. Rewritten with: maintainer/scope, status stated up top (v4.0.1-Essential released 2026-03-14 on mainline 4.0.1; nothing based on mainline 4.1/4.2 yet; last commit 2026-06-08 docs-only), downloads and build flags, a starting command, exact --speed→preset and --quality→CRF mappings from source (resolution-dependent), defaults-vs-mainline-4.2 table, tune list, and a fork-option table (zones, scd/keyint guidance, auto-tiling, enable-tf 3, enable-dlf 3, alt-cdef/dlf, distortion-bias-preset, low-memory, photon-noise, pin). Bound tables kept verbatim. Independent review: Wrong binary-limitation claim: page said release binaries lack WebM output and write IVF. The CI commit that enabled -DUSE_WEBM_IO=ON (8b8da08a, 2026-03-15 00:02Z) precedes the asset uploads (00:04-00:07Z), and the Linux_Generic binary contains libwebm/mkvmuxer and the --webm option strings. Corrected: WebM output is included and default (`--webm 0` for IVF); FFMS2, Dolby Vision and HDR10+ are not. Example output renamed to output.webm. Verified status: v4.0.1-Essential (2026-03-14) is the latest release, default branch Essential-v4.0.1, last commits 2026-06-03/08 are Parameters.md edits; build flags use-ffms2/use-webm-io/enable-dovi/enable-hdr10plus from build.sh and README. Coordinator: removed inline code formatting from a table header row, which failed WCAG colour contrast on the themed table headers. | 10/8/2026 | 7 |
| SVT-AV1-HDR | current | PassReview basis· claude-opus-5-5 · high · Correction check Snapshot moved from commit 00333404 (4.1-based September build) to the v4.2.0 tag (9dabe3ca); option code is identical at the a46bc2d PGO-build commit named in the frontmatter. Added the eight missing 4.2 options (--ivf, --obu, --cqp, --max-intra-bitrate-pct, --max-inter-bitrate-pct, --enable-kf-tf, --enable-intrabc, --enable-qmpsnr); --tune is 0–6 with 6 = Film Grain (verified: 7 is rejected); --noise-chroma is −1–200; fixed --alt-lambda-factors default (1 → 0) and --hbd-mds default (0 → −1, preset); preset range −3–13 with the 11/9 remapping; removed the stale intro paragraph about the old commit; all descriptions rewritten from source and the fork README. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The statement that the project publishes no first-party binaries is now false: the official repository released current Linux, Windows, and macOS PGO build assets on 2026-08-24. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected first-party binary availability to the September 1 PGO release and added source-scoped SDR, grain, PQ signaling, metadata and muxing workflows; synthetic native encodes decoded successfully. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks the complete fork-specific registered inventory (151 rows at source revision 00333404f455), default initialization and authored parameter explanations against pinned source. Adds practical native/pass-through combinations without borrowing mainline defaults; fork examples remain source-checked rather than newly binary-tested. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Page described the 4.1-based September 1 build and told readers to use --tune 5 for film grain; in current v4.2.0 (and builds since pgo-build-2026-09-10) film grain is tune 6 because mainline 4.2 added tune 5 = VMAF. Replaced ~1,500 words of padded prose (build-provenance narration, smoke-test narrative, license note, isolation experiments, ritual ffprobe/ffmpeg validation commands) with: maintainer/purpose, v4.2.0 and the current PGO build with asset selection order, upstream starting commands (tune 0 / tune 6 / tune 3+AVIF), a verified tune table, a defaults-vs-mainline-4.2 table, a fork-only option table, grain retention vs synthesis, and an HDR10 command. Bound parameter tables kept verbatim with a new intro stating they reflect commit 00333404 and listing what changed in 4.2. Independent review: Invented warning text: page said the encoder warns tune 6 'works best on 1080p, 4K and 8K'. The actual warning (enc_settings.c at a46bc2d) says it is optimized for moderate-to-heavy grain and will likely harm metric performance; corrected. Default GOP was 'up to 305 frames'; the code caps a 10 s GOP at 304/300 frames depending on mini-GOP size; reworded to 'capped at about 300 frames'. Removed unverified 'as in mainline' gloss on tune 1. Verified: tune 6 = Film Grain at v4.2.0 (0-6 range), tune 6 equivalence string from README, defaults (preset 4, 10-bit, QM 1/6/10, chroma QM 8, VB on, tf 1, kf-tf 1, sharpness 1, ac-bias 1.0 with sharp-tx 1, qp-scale-compress 1.0), fork-only option ranges/defaults (cdef-scaling 15, enable-qmpsnr -1, noise-adaptive-filtering 2, tx-bias modes), build flags enable-dovi/enable-hdr10plus (build.sh; README's enable-libdovi is wrong), v4.2.0 tag (2026-09-19) and pgo-build-2026-09-29-a46bc2d = v4.2.0+3 with the stated asset order. Coordinator: removed inline code formatting from a table header row, which failed WCAG colour contrast on the themed table headers. | 10/8/2026 | 6 |
| SVT-AV1-PSY | current | PassReview basis· claude-opus-5-5 · high · Correction check v3.0.2 remains the final tag. --preset is −3–13 (was −1–13) with default 8 (help says 10) and presets above 10 run as 10; --ss is −1, 0 or 1 (was '-1, 0, -1'); added the missing --tier, --intra-period and --lambda-scale-factors; --tbr default 2000 kbps; --scd is switched back off with a warning; keyint −2 is about 10 s; all descriptions rewritten. Prose fix: v3.0.2 already uses --luminance-qp-bias (--frame-luma-bias was the earlier name). Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately presents the original project as archived, separates PSYEX/HDR/mainline lineage, and limits old commands and quality claims to pinned historical reproduction. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks the complete fork-specific registered inventory (137 rows at historical 3.0.2 revision e44e0a2f6408), default initialization and authored parameter explanations against pinned source. Adds practical native/pass-through combinations without borrowing mainline defaults; fork examples remain source-checked rather than newly binary-tested. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Rewrote the page as a concise historical entry: who made PSY, its end (announced 2025-04-20 with final release v3.0.2), the archived psy-ex/svt-av1-psy repo whose README now names SVT-AV1-HDR as the official continuation (the old page's 'archived April 30, 2025' is now shown by GitHub as Feb 12, 2026, so the date was dropped). Added a verified table of which PSY features reached mainline (with versions from the mainline changelog) and which never did, a PSY v3.0.2 vs mainline v4.2.0 default/tune-number comparison, and a tested mainline command reproducing the merged PSY defaults. Removed the licence admonition, the reproducibility narrative, the 'not executed' note and stale build caveats. Bound parameter tables kept verbatim. Independent review: Verified: repository is psy-ex/svt-av1-psy (gianni-rosato URL redirects), archived=true per GitHub API; v3.0.2 'Supernova' published 2025-04-20; README end-of-project notice dated 20 April 2025 and names SVT-AV1-HDR as the official continuation. Verified PSY v3.0.2 defaults in enc_settings.c (tune 2, qm-min 2, variance-octile 6, tf-strength 1) and the mainline-port table against the mainline CHANGELOG and the SVT-AV1-HDR README's 'Merged to Mainline' list. | 10/8/2026 | 6 |
| SVT-AV1-PSYEX | current | PassReview basis· claude-opus-5-5 · high · Correction check v3.0.2-B remains the latest tag. Same preset (−3–13, above 10 runs as 10) and --ss fixes as PSY; added --tier, --intra-period and --lambda-scale-factors; fork defaults from source (tune 0, qm-min 4, chroma-qm-min 10, psy-rd 1.0, qp-scale-compress 1 on a 0.0–8.0 scale, noise-adaptive-filtering 0); --complex-hvs, --low-q-taper and --sharp-tx described from the maintainer's README (SAD→SSD, ~20% slower; avoids q below 11). Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The v3.0.2-B release, commit, lifecycle framing, named perceptual controls, lineage, and restrained evaluation advice match the fork's official material. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks the complete fork-specific registered inventory (141 rows at 3.0.2-B revision 3beb4aee471c), default initialization and authored parameter explanations against pinned source. Adds practical native/pass-through combinations without borrowing mainline defaults; fork examples remain source-checked rather than newly binary-tested. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Rewrote the page into a fork reference: status (v3.0.2-B released 2025-09-27, last commit Nov 2025, on hold per the PSY README, SVT-AV1-HDR recommended), base (mainline 3.0.2, so no 3.1–4.2 features), release binaries and build flags (enable-dovi/enable-hdr10plus; README's 'enable-libdovi' is wrong), tested starting commands, a defaults-vs-mainline table, and a fork-only options table (psy-rd, complex-hvs, spy-rd, sharp-tx, noise-norm-strength, noise-adaptive-filtering, kf-tf-strength, low-q-taper, hbd-mds with its mainline-incompatible numbering). Removed the 'Safe evaluation' checklist, the --help/--version 'save output' step and the 'source-checked, not executed' note. Bound parameter tables kept verbatim. Independent review: Wrong claim that PSYEX lacks quarter-step CRF: v3.0.2-B has extended_crf_qindex_offset (quarter steps) and CRF 1-70; corrected. Clarified that mainline's --ac-bias is absent but --psy-rd is its equivalent. Verified status: v3.0.2-B published 2025-09-27, last commit 2025-11-09, not archived; 'on hold' attributed to the PSY README ('put on the backburner'). Release assets match (Linux AVX2/AVX-512, Windows x86-64-v3). Default tune 0 and qm-min 4 / chroma 10 / psy-rd 1.0 match enc_settings.c; hbd-mds numbering matches app_config.c. | 10/8/2026 | 6 |
| SVT-HEVC | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page correctly records the explicit discontinuation and v1.5.1 final release, treats all usage as pinned legacy work, and avoids reviving stale recommendations. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Was a hedged 'historical leads' page with checksum/record-keeping advice and a Validation section but no actually usable information. Rewritten short: what it is (BSD+Patent, x86/Xeon-optimized, encMode 0–11 default 7), discontinued and now archived (last release v1.5.1, 2 June 2021), never merged into FFmpeg (plugin patches only), x265 --svt backend, a minimal legacy command, rate-control basics, and the restored stdin-pipe warning with -n. Independent review: 10-bit advice was wrong for Y4M: SvtHevcEncApp v1.5.1 takes bit depth from the Y4M header (10-bit Y4M file and pipe both produce Main 10 without -bit-depth); now says -bit-depth 10 (with -w/-h/-fps) is for raw YUV. Cut the unsourced 'x265 also compresses better' clause. Verified: repository archived with Intel discontinuation notice, last release v1.5.1 (2021-06-02), encMode 0–11 default 7, -q default 32, -rc 0/1, -tbr in bits/s, x86-only, x265 --svt option, stdin pipe with -n terminates correctly. | 10/8/2026 | 5 |
| SVT-VP9 | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The v0.3.1 status, standalone-versus-plugin boundary, supported tuning categories, conservative command shape, and experimental qualification agree with upstream. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Status stated plainly: not archived, but effectively in maintenance mode (v0.3.1 Dec 2025 is the first release since 2020; recent commits are FFmpeg-patch rebases and fixes); 8-bit 4:2:0 only. Replaced the generic controls and validation boilerplate with a tested build and stdin encode, an options table with app defaults verified in source (enc-mode 9, tune 1, QP 45, the user guide contradicts itself on QP), and the FFmpeg plugin patch's options. Added the pitfall that stdin input without -n can hang. Independent review: Status verified: v0.3.1 released 2025-12-04 (first since v0.3.0, 2020-10-30); recent commits are FFmpeg-patch rebases, formatting and a few fixes; not archived. App defaults verified in EbAppConfig.c (enc_mode 9, tune 1, qp 45, rc 0, tbr 7000000, intra_period −2); FFmpeg patch options (preset 0–9 default 9, tune vq/ssim/vmaf, rc, qp 0–51 default 32, socket) verified in the v0.3.1 patch. Missing pitfall: the library rejects widths or heights that are not multiples of 8 ('Source Height must be a multiple of 8', EbEncHandle.c). Found when a 320×180 test failed; added one sentence with a crop example. Tested the stdin pipe command with -n (24 frames decoded) using the v0.3.1 build. | 10/8/2026 | 6 |
| TapeDeck | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The default-branch status is stale: master advanced to ec360e57 on 2026-08-25 with source-path and local-scheduling hardening after the page's 5cb16dd0 snapshot, so current operational claims require revalidation. Earlier pass: 2026-09-07GPT-6 Astra · medium Updated reviewed master to August 30 commit 7578eac6, pinned documentation links, and recorded source-path/scheduling and worker metadata hardening. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Updated stale release state (page cited mixtape-v0.1.0-rc.1, commit 7578eac and Go 1.26.2; v0.1.0 was tagged 2026-10-01 and the README now states Go 1.27.0) and replaced the reviewed-state tables, architecture diagram and 'implementation claims' hedging with a short description, a capability table (encoders, audio, images, muxing, HDR/Dolby Vision modes, Vship metrics) and how deployment works, pointing to the README for exact Compose files. Independent review: v0.1.0 (2026-10-01) was preceded by mixtape-v0.1.0-rc.1, so 'first release tag' became 'first stable release'. Capability table, DV strip/keep/convert_8, port 7700, PostgreSQL stack, Arch/CachyOS script and MIT license checked against the v0.1.0 README. | 10/8/2026 | 3 |
| tar | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote with create/list/extract, the compression flag table (-z/-J/-j/--zstd/-a/-I), backup flags (--xattrs --acls --numeric-owner), reproducible archive flags and the classic -cfz pitfall. Independent review: Verified -caf for .gz/.xz/.bz2/.zst, -I brotli round trip, reproducible-archive command (identical output after touching files), -cfz creating a file named z, member extraction. Only prettier reformatting. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The draft has no primary sources and overstates portability of GNU tar syntax to BSD tar, macOS, and PowerShell while omitting format variants, metadata portability, and extraction-safety concerns. Earlier pass: 2026-09-07GPT-6 Astra · medium Scopes examples to GNU tar, separates tar variants and compression, and documents metadata portability and extraction/restore checks; archive round-trip tested. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked tar format capabilities, reference implementations, practical options and resource/integrity boundaries against primary sources. A 37,888-byte mixed binary/text fixture round-tripped exactly with the documented frontend. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Replaced generic compression/CRC wording with tar's archive-versus-filter distinction and header-only checksum scope. GNU tar 1.35 create/list/extract-to-stdout/cmp passed; metadata and filesystem restoration remain separate tests. | 10/8/2026 | 2 |
| Terminology | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Fixed wrong definitions from the import: 'bitstream' is not a media file; a 'codec' is a format, not 'the piece of code'; bitstream filters act on compressed packets, not the container. Fixed MP4 claims (AV1, VP9, Opus and FLAC have official MP4 mappings, not 'unofficial, hacky'; removed the unfounded stream-count limit and DivX/Xvid-as-codec), corrected Butteraugli as a 'video' metric, and added rate control, preset, keyframe/GOP, bit depth and chroma subsampling. Kept #container, #mp4--m4v and #discrete-cosine-transform-dct anchors used elsewhere. Independent review: Definitions (codec vs encoder, bitstream filters h264_mp4toannexb/hevc_metadata, MP4/WebM codec lists, DCT/MDCT usage, chroma subsampling) are accurate and consistent with choose-a-codec and container pages; inbound anchors #container, #mp4--m4v, #chroma-subsampling, #discrete-cosine-transform-dct intact. | 10/8/2026 | 5 |
| Theora | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Cut a long stack of timing, header and validation prose to a short legacy-codec page: what it is, one libtheora command, and browser status from MDN (Chrome 3-121, Firefox 3.5-126, Edge 12-121, never Safari). Confirmed libtheora 1.2.0 (March 2025) is the latest release and that -q:v spans 0-10. The original page was an empty stub ('To be filled'). Independent review: Browser status corrected from 'browsers are dropping it' to 'have dropped it': MDN lists support ending at Chrome/Edge 121 and Firefox 126, never Safari. History is now sourced: theora.org news gives the bitstream as published in 2004 and libtheora 1.0 on 2008-11-03. libtheora 1.2.0 was confirmed on the Xiph mirror (2025-03-29). Explained the low-quality default: without -q:v, FFmpeg's generic 200 kb/s applies (verified byte-identical with -b:v 200k). Tested the libtheora/libvorbis command and confirmed the encoder's pixel formats. Removed a generic 're-encoding adds loss' bullet. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This is an incomplete help-wanted stub whose encoding section literally says 'To be filled' and provides no capabilities, profiles, current tools, quality guidance, or sources. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked Theora 8-bit multi-chroma and fixed-header timing scope, libtheora software path, and current Mozilla support evidence. Kept legacy guidance while explicitly limiting current browser suitability and requiring tested access renditions. Locally smoke-tested FFmpeg 9.0.2/libtheora 4:2:0 Ogg output and successful software decode. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Distinguished legacy browser relevance from Xiph libtheora 1.2.0 March 2025 release archive; retained9.0.2 example scope and quoted optional audio mapping for zsh. | 10/8/2026 | 5 |
| Tools | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Coordinator: links now point to the consolidated 'Reading option tables' page instead of the retired parameter pages. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Removed the 'tested synthetic examples vs commands checked only against primary docs' narration and replaced the generic blurb with a short orientation: where to start (FFmpeg, Av1an, MKVToolNix/MP4Box, VapourSynth, players), section overview, and links to the parameter references. Independent review: Orientation page checked: all internal links resolve; no claims needing correction. | 10/8/2026 | 0 |
| Transfer Characteristics | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Rewrote with the full transfer table (FFmpeg names, usage), the fact that codes 1/6/14/15 share one curve and are displayed with BT.1886 by libplacebo-based players, a clear BT.709 vs BT.1886 explanation (the original import wrongly labelled code 1 'BT.1886'), the original's sRGB-vs-pure-2.2 community tip (tag bt470m) restored with its justification, and concise PQ/HLG and tagging-vs-tone-mapping sections. Independent review: Verified transfer code table and names; code 1 = BT.709 (codes 6/14/15 functionally identical); libplacebo v7.360.1 libav_internal.h maps bt709/smpte170m/bt2020-10/bt2020-12 to BT.1886 and bt470m to GAMMA22 (supports the gamma-2.2 tagging tip); SvtAv1EncApp spellings (bt601, srgb, hlg) and aomenc (bt2020-10bit, lin) from their help output. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Corrected the BT.709 versus BT.1886 distinction, piecewise sRGB transfer, BT.2020 SDR identifiers, and PQ/HLG relationships. Retained dated CICP identifiers and removed universal player fallback assumptions. Clarified that signal tags do not perform tone mapping. | 10/8/2026 | 8 |
| uavs3e | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Pinned the reviewed e1ff0f3 revision and distinguished the older 1.1 tag and historical patches from current source. Identified the uavs3enc executable and documented seekable raw input, rate-control units, frame limits, speed, threading and reconstruction controls. Removed a universal codec ranking; native execution was unavailable. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Previous page pinned an arbitrary commit, called v1.1 the latest tag (v1.3 exists), gave a CRF example where -q is silently ignored, and had a stale compiler note. Rewrote with a build line that actually works on current toolchains (CMake 4, GCC 16/Clang 23), correct CRF usage (--rc_type 1 --rc_crf), the real option table, and dropped the 'competitive with AV1/VVC' claim. Independent review: 'Has not changed since October 2022' was wrong: master's last commit is 2023-08-03 (LoongArch fix); tags v1.3 (2022-08-23) and v1.1 (2022-09-23). Fixed. Reran CQP and CRF commands with the author's v1.3 build; confirmed speed_level 0 is the default and slowest (16 s vs 2.4 s at level 4 on a tiny clip), rc_bitrate/rc_max_bitrate in kbps, rc_crf 0–63, -q range, BSD 4-clause license text, issue #53 still open. | 10/8/2026 | 5 |
| UT Video | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced the verification narrative with the one FFmpeg command, a note on pixel formats, and a FourCC family table. Corrected the original's outdated SSE2 requirement (current suite needs AVX2 and BMI2) and its claim of gradient prediction (the FFmpeg encoder offers none/left/median only). FFmpeg encodes only the UL* family but decodes UL, UM and UQ tags. Independent review: Misleading T2 claim fixed. The draft said FFmpeg decodes all three families and implied T2 inter-frame decoding works. FFmpeg's utvideodec.c handles the UM* tags only in packed intra mode and has no inter-frame reference handling, and the page now says so. Added the BT.601/709 FourCC pitfall: FFmpeg writes ULY* unless the stream is tagged BT.709 (-colorspace bt709 gives ULH*). Verified. Replaced the 'pass -pix_fmt yourself' advice. FFmpeg cannot encode the 10-bit UQ* family, so 10-bit sources are reduced to 8-bit; the page now points 10-bit work to FFV1. Verified the -pred values (none/left/median, no gradient) and pixel formats with ffmpeg -h encoder=utvideo. On a 10-bit 4:2:2 clip median was 7% smaller than left; none was 2.4x larger. Checked the suite version (23.2.0, 2025-06-06), the AVX2+BMI2 requirement, the GPL v2+ licence and the FourCC list against readme.en.html. Added a size/speed comparison with FFV1, checked locally on real 8-bit footage (larger files, about 2x faster). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The short page relies on Wikipedia, describes an old Windows build/format set as current, and misleadingly implies the codec generally requires SSE2 even though independent implementations such as FFmpeg have their own platform paths and capabilities. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked upstream FourCC/8-bit and 10-bit families versus FFmpeg independent encoder limits. Added exact-tag, conversion, alpha, capture-throughput and container/decode boundaries without assuming upstream decoder coverage means the same encoder supports every format. Locally smoke-tested FFmpeg 9.0.2 8-bit 4:2:2 output and successful software decode. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added suite README 23.2.0 scope, missing 8-bit UM/T2 RGB/RGBA/422/444 families and temporal compression introduced 20.0.0; explained matching tags/dependencies versus FFmpeg8-bit subset and P010/P210/v210 native 10-bit layouts. Quoted optional audio mapping for zsh. | 10/8/2026 | 4 |
| uvg266 | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page correctly distinguishes v0.8.1 from active development, uses the documented Y4M CLI shape, and appropriately qualifies bit-depth maturity, muxing, and playback support. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Replaced hedged 'validate everything' prose with a working build, three tested commands and an option table. Corrected: the build instructions changed on master (CMake rewrite April 2026), the default build is 8-bit only and rejects --input-bitdepth 10 (needs -DUVG_BIT_DEPTH=10), the default QP is 22 (not 32), and there is no CRF mode. Independent review: Found a silent-corruption pitfall: a -DUVG_BIT_DEPTH=10 build fed 10-bit Y4M without --input-bitdepth 10 encodes garbage (7 dB PSNR vs 42 dB with the flag); added that uvg266 ignores the Y4M header bit depth. Verified: v0.8.1 (2024-07-01) is the last tag, master rewrote the build system in April 2026, BSD-3-Clause, README 'still under development', defaults (qp 22, period 64, ref 4, gop by preset), UVG_BIT_DEPTH option 8/10 default 8; reran 8-bit file/pipe commands and the 8-bit build's usage error with --input-bitdepth 10. | 10/8/2026 | 3 |
| VAAPI | current | PassReview basis· claude-opus-5-5 · high · Correction check Added descriptions for all 12 av1_vaapi options, including the actual auto rate-control selection order from vaapi_encode.c and idr_interval semantics (non-key I-frames between key frames). Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks all 12 FFmpeg 9.0.2 AV1 private option spellings, ranges/named values and wrapper defaults, plus codec-specific generic initialization overrides. Hardware recipes are independently source/help-checked, not GPU-executed; driver negotiation, automatic settings and compatibility remain explicit limits. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Page covered only the AV1 wrapper with CQP/VBR/CBR experiments, a 'hardware encoding was not executed' disclaimer and a fixed -g/-bufsize recipe. Now explains what VAAPI is and which GPUs use it, how to read vainfo (EncSlice vs EncSliceLP), the simple -global_quality command (FFmpeg tries ICQ then CQP; QVBR if combined with -b:v), AV1 needing explicit -rc_mode CQP because ICQ clips to 1-51, a full-GPU hwaccel pipeline, options table and pitfalls. Title kept as 'VAAPI AV1 encoding' per brief. Marked block moved verbatim to the appendix. Independent review: ICQ -> CQP fallback verified in vaapi_encode_init_rate_control; confirmed Mesa's VA frontend advertises only CQP/CBR/VBR (+QVBR on some), so AMD falls back to CQP. AV1 ICQ clip to 1-51 (av_clip(rc_quality, 1, 51)) verified, so the explicit -rc_mode CQP advice for AV1 stands. Added the missing AVBR step: a bare -b:v with no -bufsize tries AVBR first (ignores buffer constraints). -idr_interval corrected (counts I-frames: only every Nth I-frame is an IDR); -compression_level described as the driver quality level instead of an Intel-only claim; AV1 hardware line now includes Phoenix APUs, consistent with the AMF page. vainfo EncSlice/EncSliceLP explanation and all four commands checked (parse in 9.0.2; fail only at 'No VA display found'). | 10/8/2026 | 4 |
| VapourSynth | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The R79 build requirements are materially understated: the tagged Meson build selects C++20, and Windows requires Python 3.14+, while the page says C++17 and a general Python 3.12 baseline. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected the R79 source-build requirement from C++17 to C++20 and specified Python 3.14+ on Windows versus 3.12+ elsewhere; distinguished wheel compatibility from compiling the Windows core and repaired the native encoder pipe. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked R80/API R4.3 VSPipe selection, inclusive range, Y4M and concurrency controls. A 24-frame 10-bit Y4M to FFV1 pipeline and full decode passed. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Replaced the R79/R80 'reviewed scope' tables, reproducibility checklists, license notes and the long Av1an caveat (now one sentence pointing to the Av1an page, which carries the accurate R73+ note) with practical content: install (pip, vsjetpack extras and vs-wheels, VSRepo), previewers (VSView replaces archived vs-preview), source filters (BestSource/L-SMASH-Works/ffms2 with real caveats restored from the original), piping to x264/x265/SVT-AV1/FFmpeg and a trimmed vspipe option table. Updated to R81 and noted havsfunc r34 dropped nearly all functions. Independent review: Added a caution: vapoursynth-mvutensils 10 (PyPI, 2026-10-08) breaks vs-jetpack 2.2.6's QTempGaussMC and mc_degrain ('AnalyseMany: Function does not take argument(s) named fields'); reproduced with 9 vs 10 in a copied venv. vsjetpack[full] requires mvutensils>=6 unpinned, so a fresh install hits it; pin <10. Install wording follows the official docs (pip install vapoursynth, then vapoursynth config). Verified R81 (2026-10-04) and vsjetpack 2.2.6 (2026-10-04) are current, havsfunc r34 notes, vs-preview archived, VSView install, BestSource cache paths/VC-1 note from the R22 README, all vspipe pipe commands run. | 10/8/2026 | 7 |
| VC-1 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced the preservation checklists with a profile and FourCC table, who can decode it, and two short commands (lossless remux and a one-step transcode for playback). Stated plainly that FFmpeg has no VC-1 encoder (wmv1 and wmv2 are different codecs). The original was a stub with 'To be filled'. Independent review: Cut the incorrect 'the only encoders were Microsoft's Windows Media tools' (commercial encoders existed) and kept the verifiable 'no open-source VC-1 encoder'. ffmpeg -encoders lists only wmv1/wmv2; vc1/wmv3 are decode-only, with qsv/cuvid/v4l2m2m wrappers. Sourced the Blu-ray mandatory-codec claim to Microsoft's 2004 announcement and noted the WMV9 lineage. HD DVD's identical trio is corroborated by 2004 DVD Forum coverage. Toned down 'interlaced material is common on Blu-ray' to a factual note: Advanced profile can carry interlaced video such as 1080i discs. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This is an unfinished stub and its claim that VC-1 can be decoded by any Blu-ray-capable device is overbroad; Blu-ray device support depends on disc profile, elementary-stream constraints, and carriage rather than a universal arbitrary-file decoder. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked SMPTE profile/sample-format scope, Microsoft WMV3 versus WVC1 identifiers and independent FFmpeg decoding. Avoided inventing a VC-1 encoder wrapper; retained preservation/access-copy workflow and separated legacy desktop support from browser support. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Retained already-complete profile, sample, legacy decode/remux and delivery boundaries; quoted optional audio mapping so the existing copy command works in zsh without pathname expansion. | 10/8/2026 | 4 |
| Video artifacts | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Removed the duplicate H1, tightened each artifact description and corrected causes: moiré and staircase/aliasing usually come from sampling/scaling rather than the encoder; ringing explained via quantized high frequencies; mosquito noise as its temporal counterpart; banding fixes (10-bit, AQ, grain/dither, deband). Added blur/detail loss (the most common modern artifact) and keyframe/frame-type pulsing. All images kept. Independent review: Removed the claim that encoders 'often quantize color more coarsely' (H.264's chroma QP mapping actually quantizes chroma finer at high QP); color bleed now attributed to subsampling plus low bitrate. Fixed the SVT-AV1 option name to --qp-scale-compress-strength (its documented purpose is consistency across temporal layers, matching the pulsing section). | 10/8/2026 | 3 |
| Video codec support in browsers and devices | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-haiku-5-5): Replaced the ten repeated 'No native support promise is established here' cells with shorter, consistent wording, cut the 'records observations checked 2026-10-04' and 'laboratory certification' disclaimers, and rewrote the Safari/Firefox paragraph into a single platform-check instruction. Kept the four-path table, the probe code, and the device-validation steps (the reproducibility record was reduced to one line). Frontmatter sources, lastVerified and tags unchanged; description rewritten. Independent review: Replaced the defensive 'what documentation establishes' table (ten 'No native support documented in these sources' rows) with a direct native-playback table; the ten unsupported formats are one row that still links each codec page. Theora fixed: Firefox 126 disabled it by default (MDN Firefox 126 notes) and Chromium's codec list no longer includes it. Chrome HEVC needs hardware support (Chromium list); WebM VP8/VP9 on macOS since Safari 14.1 and iOS/iPadOS since 17.4 verified in WebKit's 17.4 notes. NVIDIA matrix URL updated to the current page. Coordinator: reworded the WebCodecs hardwareAcceleration note to state what prefer-hardware means instead of a disclaimer. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Verified primary Chromium, Mozilla and WebKit browser observations dated October 4, 2026, API roles from W3C specifications and hardware encode/decode distinctions from vendor documentation. JavaScript examples are illustrative probes; no unperformed device laboratory certification is claimed. | 10/8/2026 | 10 |
| Video codecs overview | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the 16-row, five-column hedged comparison table (with unverified cells such as 'x266 availability unverified', DSV2/ProRes precision claims) and the 'Compatibility claims need scope' caution with two short tables: delivery codecs (standardiser/year, use today, main encoders) and production/archive/legacy codecs. Added a practical 'picking a codec' checklist and one ffprobe command to identify codec/profile/level/pix_fmt. Independent review: Codec chooser checked against /guides/choose-a-codec and /reference/video-codec-support: they agree on the AV1 + H.264 fallback ladder, VP9 as a web fallback, HEVC for TVs/Apple/HDR and VVC for known targets only. DSV/DSV2 row reworded to 'Hobby wavelet codecs; experimentation only' to match those pages' descriptions. All internal links resolve; the ffprobe identification command was run on a synthetic H.264 MP4. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The overview gives a useful and appropriately qualified comparison of codec maturity, implementations, containers, HDR signaling, and whole-delivery-stack constraints. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Verified the overview links every listed video codec to its expanded reference and support guide, scopes archived/experimental encoder entries, and differentiates delivery, editing, preservation and research formats. Codec ranking remains explicitly dependent on implementations and workload. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Restored the complete 16-codec comparison as one contiguous GFM table by removing the blank line after FFV1; parsed generated Markdown HTML to confirm all rows and five columns, and formatted the entire owned table. | 10/8/2026 | 4 |
| Video Players | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Replaced the stale MPC-HC 2.8.1 version and disclaimer-heavy 'validate the presentation' prose with a player table (mpv, VLC 3.0.24 with 4.0 still unreleased, MPC-HC clsid2 2.8.3, MPC-BE 1.9.1, IINA 1.5.0, mpv front ends), mpv as the tool for checking encodes (software decoding is mpv's default, so the redundant --hwdec=no was dropped), verified default key bindings for frame stepping/exact seeking/stats/screenshots, a --lavfi-complex hstack side-by-side command from the mpv manual, and an HDR note based on gpu-next being the default VO in 0.41 and --target-colorspace-hint. Independent review: Verified versions (mpv 0.41.0, VLC 3.0.24, clsid2 MPC-HC 2.8.3 released 2026-10-08, MPC-BE 1.9.1, IINA 1.5.0) via GitHub releases/videolan.org. Verified default key bindings (., ',', Shift+arrows exact seek, i/I stats, s/S screenshots, #/j, [/], Ctrl+h hwdec toggle), JPEG screenshot default and --target-colorspace-hint (Wayland/D3D11/winvk, auto) against v0.41.0 input.conf and options.rst. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This incomplete recommendation list contains a TODO, no reviewed versions or capability evidence, and broad compatibility claims that are not tied to current mpv, VLC, or MPC-HC releases. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced incomplete universal recommendations with release-scoped player roles, concrete mpv diagnostics and presentation validation; no cross-platform matrix is claimed. | 10/8/2026 | 7 |
| VideoToolbox | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced the 'AV1 availability' hedging and the SDR baseline with a direct page: encoders (h264/hevc/prores_videotoolbox, no AV1; Apple's AV1 hardware is decode only), -q:v 0-100 constant quality on Apple Silicon only (source: vtenc_qscale_enabled), -tag:v hvc1, 10-bit via -profile main10 -pix_fmt p010le, Intel Mac -b:v fallback, scale_vt full pipeline, options table, brief HandBrake note. Removed redundant -allow_sw 0 (default is already 0). Independent review: -q:v condition made precise: vtenc_qscale_enabled() is !TARGET_OS_IPHONE && TARGET_CPU_ARM64, i.e. native arm64 macOS builds only; added that an x86_64 FFmpeg under Rosetta also lacks it. ProRes hardware requirement corrected: base M2 (and later) chips have the ProRes engine per Apple's M2 announcement, not only M1 Pro/Max/Ultra. HandBrakeCLI encoder names verified in libhb/common.c (vt_h264, vt_h265, vt_h265_10bit) and vt_prores added; HandBrake's 'VideoToolbox scaler is less sharp' confirmed in its docs. No AV1 VideoToolbox encoder in FFmpeg 9.0.2 confirmed (allcodecs.c). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The entry declares itself incomplete, uses literal [0-100] placeholders in copyable commands, and makes broad model/year codec claims without teaching users to query the encoders and properties actually available on the host. Earlier pass: 2026-09-07GPT-6 Astra · medium Removed the incomplete notice and literal quality placeholders; added host encoder/session-property discovery, a numeric H.264 bitrate example with software fallback disabled, and codec/HDR checks tied to the actual Mac. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks that FFmpeg 9.0.2 source registers no av1_videotoolbox encoder, without generalizing that absence to all Apple APIs or future versions. Preserves separately scoped H.264 guidance and explicit no-hardware-execution limits. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. | 10/8/2026 | 5 |
| VMAF | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Verified the previously suspicious claims: VMAF v1 (June 2026, libvmaf 3.2.0), model/vmaf_v1.0.16/vmaf_v1.0.16_3d0h.json, the 0-110 4K 3H model, HFR variants and the Speed_chroma issue (3.2.0 only compiles speed_chroma with -Denable_float; fixed in 3.2.1) are all real, so kept and explained them. Added the community stance (enhancement gain/NEG, poor discrimination near transparency with KonFiG-IQA data, v0 luma-only, no HDR model, viewing-condition dependence), a model table (v0.6.1, NEG, 4K, phone via enable_transform, all v1 built-ins), minimal and complete FFmpeg commands (10-bit for v1, n_threads, JSON log, NEG side by side, bicubic upscale, worst-frame jq), standalone vmaf CLI with Y4M and named pipes, and a condensed Windows model-path tip tested on Windows. Independent review: VMAF v1 independently verified as real: libvmaf v3.2.0 released 2026-06-20 'Initial public release for VMAF v1'; v3.2.1 (2026-09-14) 'enable speed_chroma/speed_temporal without enable_float'; models_v1.md lists vmaf_v1.0.16_3d0h/5d0h/1d5h_2160/3d0h_2160 with 3d0h_2160 at [0,110]; model/vmaf_v1.0.16_hfr/ has the four _hfr variants; libvmaf/src/model.c at v3.2.1 registers all eight as built-in versions; v1 model JSON features are CAMBI, Speed_chroma, ADM3 (adm_enhn_gain_limit 1.0 = NEG) and motion3, no VIF. Opening paragraph now tells newcomers why a VMAF number alone can't rank encoders. Added a concrete enhancement-gain example verified in source: SVT-AV1 4.2.0 TUNE_VMAF = 5 applies luma unsharp masking (definitions.h, enc_handle.c warning), libaom --tune=vmaf searches an unsharp amount that maximises VMAF (tune_vmaf.c). Did not use 'vmaf_with_preprocessing', which no longer exists in libaom main's arg_defs.c. Added Netflix's own statement (v1 tech blog) that film-grain noise and perceptual optimisations such as adaptive quantisation remain unsolved; HDR-v1-planned and VIF-removal claims confirmed from the same post. Named-pipe vmaf example hung in an interactive shell: both background FFmpeg jobs show 'Stopped (tty output)'. Added -nostdin (verified working under a pty) and explained it; -strict -1 still required by FFmpeg 9.0.2 for 10-bit Y4M. Added v1 CAMBI encode-side parameters (cambi.enc_width/height/bitdepth) from models_v1.md; tested with the vmaf CLI using version= (feature key shows encbd/ench/encw). Changed '4K 1.5H = Netflix's default' to 'the default choice for 4K' (models_v1.md wording); added AIM term to the v1 feature list. Tested: default libvmaf; v1 10-bit with JSON + n_threads; two models with name= (vmaf 78.22 / vmaf_neg 75.98); enable_transform phone mode (93.6 vs 78.2); pool=harmonic_mean; jq worst-frames query; vmaf CLI on 10-bit Y4M matches the FFmpeg filter exactly (78.799189). Coordinator: removed inline code formatting from a table header row, which failed WCAG colour contrast on the themed table headers.
Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page contains broken commands and typos such as mkfido, stale build and model guidance, and oversimplified scoring and temporal claims; it also predates libvmaf 3.1 and Netflix's 2026 v1 model set, so it is unsafe as a current usage reference. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaces broken legacy usage with explicit model selection, v1/SDR/viewing-condition guidance and v0 reproducibility. FFmpeg v0 plus standalone/FFmpeg v1 smoke-tested; v1 required matching source-built extractors, not merely the installed 3.2 version label. | 10/8/2026 | 10 |
| Vorbis | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced capability table and "verify" prose with: Opus supersedes it, libvorbis -q:a 5 / oggenc -q 5 commands, quality scale, not to use FFmpeg native vorbis, Safari 18.4 Ogg note, pitfalls. Restored original context on Minecraft use. Independent review: 'Released in 2000' corrected to Vorbis I finalised in 2002; 'Opus is better at every bitrate' softened to 'more efficient, especially at low bitrates'. Verified FFmpeg libvorbis default q 3, -q range -1..10 (oggenc scale), -iblock -15..0 default 0 from libvorbisenc.c; libvorbis -q:a 3/5/7 commands run. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The very short page is unsourced, contains a broken/truncated description and spelling errors, and reduces current selection guidance to an unqualified claim that Opus has largely replaced Vorbis. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds specification-backed header rates, 1–255 signaled channels, ordering and coding-mode distinctions. Checks the single FFmpeg private iblock option/default and VBR setup against source/help, executes the basic example, and updates Safari Ogg guidance from the vendor release. | 10/8/2026 | 5 |
| VP8 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Cut from ~2,500 to ~1,000 words: condensed the long coding-tools and decoding-process narrative, merged the duplicated capability summary into one format table, and kept the community history (MSU comparison image, Fiona Glaser's analysis) in a short section, now tied to RFC 6386's statement that the reference source defines the bitstream. Fixed the version-field description (versions 0-3 select interpolation/loop filters per RFC 6386), added WebRTC's VP8 mandate (RFC 7742), alpha carriage and the 2013 Google/MPEG LA agreement. Replaced the padded 4-line command block with one tested command and a correct explanation of VP8 constrained quality (FFmpeg falls back to 256 kbit/s without -b:v). Added a RelationshipTable for implementations. Independent review: Version field was described as selecting loop filters; RFC 6386 says the version's loop-filter setting has no effect on decoding and only the reconstruction (interpolation) filter changes. Fixed. History verified: RFC 6386 (Nov 2011) states the bitstream is defined by the reference source code, not the narrative. Rate control re-tested: -crf without -b:v logs 'using default of 256kbit/sec' and gives q=58; replaced the placeholder disclaimer with that concrete behaviour. Removed claim that VP9/AV1 are 'just as widely supported in browsers' (AV1 is not in all Safari). Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked RFC 6386 8-bit 4:2:0 coding and version semantics, VP8 RTP carriage and libvpx workflow. Distinguished WebP lossless from VP8 video and native file/MSE/browser support from realtime decoding; retained the technical reference with targeted wording repairs. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Repaired truncated description and fixed-PSNR visual guarantees, corrected base quantizer index/component list and 16-bit arithmetic overclaim, removed TrueMotion uniqueness/frequency claim, corrected MV coded-difference units/range and versioned interpolation, labeled old benchmarks historical, added locally tested FFmpeg 9.0.2 VP8/Opus constrained trial and quoted optional map. | 10/8/2026 | 7 |
| VP9 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Rewrote around use today (YouTube/WebRTC, AV1 fallback, web HDR), a single profile table, plain-language coding tools (alt-ref frames, segmentation, reference scaling), encoders and hardware encode reality (Intel only; no NVENC/AMF VP9), one tested FFmpeg command with corrected rate-control explanation (-crf alone is constant quality in current FFmpeg; -row-mt is off by default in libvpx v1.17.0), decoder/hardware generations, containers and the HDR detail that VP9 has no transfer-function field so HDR relies on container metadata (vpcC + SmDm/CoLL or Matroska Colour). Removed the six-step workflow, the duplicated profile discussion and the hedged licensing note. Independent review: 'Only Intel GPUs encode VP9' was wrong for mobile: WebRTC's Android HardwareVideoEncoderFactory uses Qualcomm and Exynos MediaCodec VP9 encoders. Reworded to 'on PCs only Intel', plus the Android SoC note. NVIDIA decode history corrected: 'second-generation Maxwell' was too broad (GTX 970/980 have no VP9 decode); now GTX 950/960 8-bit, Pascal 10/12-bit. Verified libvpx v1.17.0 is latest and row_mt defaults to 0; FFmpeg libvpxenc selects VPX_Q for VP9 when -crf has no bitrate. Command tested (Profile 0 output). Safari 14.1/17.4 milestones match /reference/video-codec-support.
Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page correctly presents VP9 profiles, pixel formats, WebM and ISOBMFF carriage, color/HDR limitations, current implementations, and reproducible evaluation guidance without unsupported efficiency promises. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked WebM profile/depth/chroma and level tables, libvpx workflow, lossless input-sample limits, alpha carriage and decoder/hardware boundaries. Added dated support-guide integration rather than claiming all browser profiles and containers work. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Corrected omitted 4:4:0 in profiles 1/3 using v0.7 spec sections 7.2/7.2.2 and tagged libvpx decoder; distinguished the shorter WebM summary and actual encoder/hardware coverage; updated stale v1.15.2 scope to directly checked v1.17.0 tagged source. | 10/8/2026 | 11 |
| vpxenc / libvpx | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrote all 111 native rows against vpxenc.c and vp9_cx_iface.c at v1.17.0; defaults matched, behaviour was fixed: VP9 --kf-min-dist only 0 or equal to --kf-max-dist; --timebase only with --fps; --disable-warnings does nothing in 1.17.0; --resize-up/down and --output-partitions are VP8-only; VP9 --noise-sensitivity is 0–6, realtime-only and ignored above 8-bit; alt-ref layers (auto-alt-ref >1) only in pass 2; --cpu-used rules; 13 short aliases added. FFmpeg libvpx-vp9 table gained a meaning column. Key finding: the legacy alias defaults win, so FFmpeg sends arnr-maxframes 0 / arnr-strength 3 / lag 25, i.e. alt-ref filtering is off unless set; -noise-sensitivity is never passed to VP9; -overshoot-pct >100 fails to open. Help sha256 updated to the local 1.17.0 output. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Coordinator: FFmpeg's libvpx-vp9 sends arnr-maxframes 0 (legacy option default applied last, verified in libvpxenc.c n9.0.2), so the options table and FFmpeg section now give FFmpeg's defaults and the flags to match vpxenc; tested locally. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The reviewed-version claim is stale: the official changelog lists libvpx v1.16.0 from 2026-01-06, newer than the page's v1.15.2, with ABI and feature changes relevant to builds and guidance. Earlier pass: 2026-09-07GPT-6 Astra · medium Updated to libvpx 1.17.0 and documented native/wrapper quality, two-pass bitrate, high-bit-depth and validation workflows; native 8/10-bit and two-pass samples decoded successfully. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks all 111 VP9-relevant native libvpx 1.17.0 names and 36 FFmpeg-private controls; excludes VP8-only controls and distinguishes clipped tile defaults, native two-pass behavior and wrapper sentinels. Eight encode/mux recipes passed complete decode and frame-count checks. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted two optional FFmpeg -map values so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Rewrote the guide portion as a clean libvpx 1.17.0 VP9 reference: vpxenc constant-quality (two passes by default), 10-bit profile 2, FFmpeg pipe, two-pass VBR, FFmpeg libvpx-vp9 one- and two-pass CRF, an options table and pitfalls, plus a brief VP8 section. Restored useful community guidance from the original (row-mt, auto-alt-ref 6, lower arnr-strength for grain, tune-content, kf-max-dist) and corrected original errors: tile-columns defaults to 6 (clamped to max) not 0; 2160p allows 8 tile columns not 16; multi-layer alt-ref only works in two-pass, with no 8-bit incompatibility; Y4M input doesn't need --enable-libyuv; --input-bit-depth isn't needed with Y4M. Documented the silent broken-output pitfall when piping into vpxenc VP9 without --passes=1, the 128-frame default GOP, vpxenc's default 256 kbit/s VBR, and FFmpeg -crf with nonzero -b:v meaning constrained quality. Removed the testing narrative and padded flags (-threads 4, -map, -an). Bound reference tables moved to the end verbatim, with shortened intros. Independent review: Verified against libvpx v1.17.0 (latest tag): tile_columns default 6, VP9 vpxenc passes default 2 except realtime, VBR 256 kbit/s default, kf_max_dist 128, lag max 25, g_threads 8, arnr 7/5, auto-alt-ref 1, --validate-hbd-input in the 1.17.0 CHANGELOG. 256-px minimum tile width gives 4 columns at 1080p and 8 at 2160p, consistent with Google's VOD table. Verified FFmpeg n9.0.2 libvpxenc: defaults b=0 and g=-1 (so libvpx's 128), VP9 -crf without bitrate → VPX_Q, with bitrate → CQ, VP8 always CQ, cpu-used default 1 (VP8 range −16..16). Google VOD guidance values (CRF 31/24/15, -g 240) confirmed. Tested: vpxenc CQ, 10-bit profile 2, pipe with --passes=1 (24 frames) and without (110-byte unreadable WebM, exit 0), two-pass VBR, VP8 vpxenc; FFmpeg one-pass and two-pass CRF, -crf without -b:v, 10-bit (profile 2 automatic), VP8 with cap. No changes needed. | 10/8/2026 | 6 |
| Vship and FFVship | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh Upstream released v5.1.1 on 2026-08-25, so the page's statement that v5.1.0 is latest and 5.1.1 exists only ahead on main is now false. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrected release to v5.1.1 and documented color/metric fixes, reproducibility implications and the upstream source/reference CLI documentation inconsistency. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Updated from 5.1.1 to the current 5.1.2 (6 Oct 2026), removed the interactive-chart/dataset aside, the 'not executed' narrative and stacked cautions. Clarified the --source/--reference doc inconsistency from the actual parser (--source/-s, --encoded/-e; -e is not --every), real defaults (gpu-threads 3, threads 1, qnorm 2, intensity 203, standard_fhd), build/install steps including the VapourSynth plugin symlink, backend trade-offs from release notes, the VapourSynth function/property table, and a one-sentence licence note (MIT NON-AI). Independent review: Added --verbose to the FFVship table and a short paragraph on untagged files (FFVship assumed BT.470BG matrix, transfer and primaries for an untagged 640x360 file and scored it -25.7 vs -14.37 when tagged). Verified 5.1.2 release date and assets (FFVship_nvidia/FFVship_Vulkan, libvship NVIDIA/Vulkan), 5.1.1 fixes (full-range YUV, CVVDP), backend benchmark claims, parser defaults (gpu_threads 3, cpu_threads 1, intensity 203, -e = --encoded), Makefile PREFIX default /usr/local and plugin symlink to $PREFIX/lib/vapoursynth, MITIGATE_MALLOC_ASYNC, MIT NON-AI LICENSE vs README 'MIT'. Tested FFVship SSIMULACRA2/Butteraugli/CVVDP runs and the VapourSynth plugin functions. | 10/8/2026 | 6 |
| VTM | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This Help Wanted page is only a short historical paragraph and offers no VTM version, build, configuration, supported-format, or validation information. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the Help Wanted paragraph with VTM-24.0 build instructions, a random-access ten-bit raw-input example, profile/configuration limits, and matching reference-decoder validation. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Cut to what VTM is for, why it is not a practical encoder (single-threaded; VVenC medium is ~280x faster at similar PSNR BD-rate per Fraunhofer), build for VTM-24.0 (still the latest tag, April 2026) and one raw-YUV encode/decode example. Removed the validation checklist, 'did not compile' note and license notes. Independent review: Built VTM-24.0 (source from the Software Heritage vault, tree hash identical to the tag; vcgit clone fails from this host). The documented build fails on GCC 16.2 (-Werror on unused-but-set-variable in SEIFilmGrainSynthesizer.cpp); added the tested -DCMAKE_CXX_FLAGS=-w workaround. 'VTM reads raw planar YUV only (no Y4M)' was wrong: VTM-24.0 reads .y4m and takes size, rate and bit depth from the header. Encode command rewritten to a tested Y4M form with ./bin/ paths (binaries are not on PATH); raw-YUV flags described in one sentence; decoder output is byte-identical to recon. Added that the shipped configs code at 10 bits, so recon/decoded output is 10-bit even for 8-bit input; replaced 'useful range 22–37' with the JVET CTC QPs 22/27/32/37. Removed overclaim 'Every VVC encoder and decoder is checked against it'; VVenC speed comparison now matches the wiki table (medium, 8 threads, ~280× HD to ~380× UHD vs VTM-19.2). Verified options in EncAppCfg.cpp, IntraPeriod -32 rule, RateControl/TargetBitrate default off, no threading (VVenC wiki: not supported for HM/VTM), VTM-24.0 (2026-04-08) latest tag. | 10/8/2026 | 5 |
| Vulkan Video | current | PassReview basis· claude-opus-5-5 · high · Correction check Added descriptions for all 13 av1_vulkan options: rc_mode auto uses CQP with -qp/-global_quality, VBR (else CBR) with -b:v, and otherwise fixed QP 18; -quality above the driver maximum is rejected; units hdr writes mastering-display and CLL metadata. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks all 13 FFmpeg 9.0.2 AV1 private option spellings, ranges/named values and wrapper defaults, plus codec-specific generic initialization overrides. Hardware recipes are independently source/help-checked, not GPU-executed; driver negotiation, automatic settings and compatibility remain explicit limits. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): Replaced the AV1-only experiments and 'not executed' disclaimers. States what Vulkan Video encode is, FFmpeg's status (h264/hevc_vulkan since 7.1, av1_vulkan since 8.0, per Changelog), that it is the least settled hardware path, how to check driver support (VK_KHR_video_encode_* extensions), CQP/VBR commands, a Vulkan decode+scale+encode pipeline, options table and pitfalls. Marked block moved verbatim to the appendix. Title kept per brief. Independent review: Milestones verified from allcodecs.c at tags: h264/hevc_vulkan first in n7.1, av1_vulkan in n8.0. Removed ritual '-rc_mode cqp': init_rc() selects CQP automatically when -qp or -global_quality is set and VBR (else CBR) when -b:v is set. Added the real pitfall: with no rate options FFmpeg uses a fixed QP 18 for every codec, which is huge for AV1. Build requirement completed: av1_vulkan needs Vulkan headers >= 1.4.317 (vulkan_1_4 dep), general Vulkan >= 1.3.277. bf=2/g=300 defaults apply to all three wrappers, not just AV1. Dropped the redundant -init_hw_device from the hwaccel pipeline and a stacked HDR disclaimer in the appendix intro. | 10/8/2026 | 4 |
| VVC / H.266 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Rewrote with a clear verdict (broadcast systems such as Brazil TV 3.0 and evaluation; not general distribution), profile table extended with the 2022 12/16-bit profiles, plain-language list of VVC's main coding tools, encoders/decoders with verified FFmpeg milestones (native decoder 7.0, stable 7.1, libvvenc and QSV 7.1, VA-API and Matroska 8.0), hardware decode status (Intel Lunar Lake, TV/STB SoCs), containers, HDR and licensing (VVC Advance, VCL Advance). Restored the original community observation that VVenC is roughly competitive with the best AV1 encoders. Removed the evaluation checklist and repeated caveats. Independent review: FFmpeg decoder milestone corrected: 7.0 shipped it with AV_CODEC_CAP_EXPERIMENTAL, 7.1 removed the flag (no 'declared stable' wording upstream). libvvenc and QSV VVC decode in 7.1, VA-API and Matroska in 8.0 confirmed in Changelog. Hardware decode: Lunar Lake confirmed; Intel states discrete Arc B-series (Battlemage) has no VVC decode, so that is now explicit. Cut the unverified 'a few phone SoCs'; NVIDIA/AMD have none (NVDEC matrix has no VVC). Cut unverifiable licensing claims ('several large patent holders license outside any pool', cost uncertainty as cause of slow adoption). Release versions confirmed: VVenC v1.14.0, VVdeC v3.2.1 (2026-10-07), uvg266 v0.8.1; ITU-T H.266 current edition 01/2026.
Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page is current with H.266 version 4 from January 2026 and accurately qualifies profiles, container mappings, HDR signaling, implementation maturity, evaluation, and licensing. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked current H.266 profile families, VVenC implementation documentation and VVdeC Main10 conformance scope. Distinguished native browser playback from a separately integrated WASM decoder and retained edition/container/evaluation constraints. | 10/8/2026 | 9 |
| VVenC | current | PassReview basis· claude-opus-5-5 · high · Correction check Built VVenC v1.14.0 to compare --fullhelp output (hashes matched) and rewrote 49 vvencapp rows, the vvencFFapp rows (378 options; the 64 identical --FrameN rows are now one row bound via tableName) and 9 libvvenc rows. Fixed: --stats is an on/off progress switch, not a stats path; --SAO default is 1 (help prints 0, every preset sets 1); expert defaults are preset medium's tools (vvenc_config_default applies MEDIUM) with QPA off and Threads 0; truncated '[' list defaults → none; --format and --PackedInput cells were garbled help fragments; threads auto rule; ranges for Affine/MMVD/AMVR/Geo/etc. from the parameter checks; short aliases added. libvvenc gained a description column plus how the wrapper maps -b:v, -g, -pass, colour tags. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Coordinator: the vvencFFapp starting command now enables --PerceptQPA=1, which the expert app leaves off. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The v1.14.0 release, vvencapp/vvencFFapp/library split, input format, option examples, FFmpeg integration, and interoperability warnings align with official sources. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks VVenC 1.14.0 simple versus expert/config defaults, including hidden GOP entries and aliases, plus nine source-only FFmpeg libvvenc controls. Six native recipes passed strict full decode/frame counts; lossless sample matched input, while absent wrapper execution is not claimed. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): The page opened with a build/--version/--fullhelp ritual and ~8,000 words of option tables before any usable command. Replaced the top with a short guide (interfaces, build, starting commands for vvencapp/FFmpeg/vvencFFapp, presets, rate control, key options, muxing and decoder support) and moved the three validator-bound reference sections verbatim under '## Full option reference'. Removed the 'exercised on 2026-10-01' narrative. Independent review: Expert-app command was wrong in two ways: cfg/sequence.cfg forces Level 4.1 and FramesToBeEncoded 600 (not just placeholders), and vvencFFapp defaults to Threads 0 (single-threaded). Replaced with 'vvencFFapp -c cfg/randomaccess_medium.cfg -i input.y4m -t -1 -q 32 -b output.266' (tested; vvencFFapp reads Y4M). --additional example 'MaxPicSize=3840x2160' fails ('if max picture size is set, both RPR and resChangeInClvsEnabled have to be enabled'); replaced with tested 'MCTF=0:BIM=0' (MCTF=0 alone fails because BIM requires MCTF). Matroska: raw .266 → MKV fails in FFmpeg 9.0.2 even with -framerate/-fflags +genpts, and mkvmerge 102 does not recognise raw VVC; added one sentence recommending MP4. Speed claim made precise per the Encoder-Performance wiki Table III (medium, 8 threads, ~280× VTM-19.2 on HD, ~380× on UHD). Verified: v1.14.0 (2026-01-21) latest, VVdeC v3.2.1 (2026-10-07), BSD-3-Clause-Clear, vvencapp help (preset list incl. medium_lowDecEnergy, -p default two-pass, -m ≥1.5× target enforced, thread defaults by resolution, -rs 1, --sdr/--hdr values), libvvenc wrapper (yuv420p10le only, -qp default 32, -period, -flags +cgop → IDR, -vvenc-params), FFmpeg changelog VVC entries (7.0 experimental, 7.1 libvvenc + QSV, 8.0 VAAPI and VVC in Matroska), and that a later --preset resets earlier expert options. | 10/8/2026 | 9 |
| WavPack | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote around what the native tools do (verified by building WavPack 5.9.0): lossless default, -h/-hh/-x, hybrid -b/-c with the bits-per-sample caveat, FFmpeg encoder is lossless-only and ignores .wvc. Restored the original import's practical notes on hybrid mode, FLAC comparison and adoption. Independent review: Cut unverified 'compresses a little better than FLAC' (now 'about as well as FLAC') and the invented '1.5x and 2x slower' multipliers for -h/-hh. Verified against the WavPack 5.9.0 manual: -b ranges (2.0-23.9 bits/sample or 24-9600 kbps, ~196 kbps floor for CD audio), -c/-cc, -x0..6 default 1, --threads up to 12 and not for hybrid stereo, float and DSD support. FFmpeg wavpack encoder compression_level up to 8 (wavpackenc.c). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh Although much of the technical outline is useful, important recommendations and adoption claims are unsourced and overbroad, including the claim that WavPack offers little benefit without hybrid support and the speculative DRM rationale for FLAC-in-MP4 adoption. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaced the failed draft with sourced capabilities, practical workflow, explicit implementation limits, and output validation; human editorial approval remains pending. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds current native-project integer, float, DSD, channel/rate and metadata capabilities, distinguishing modern features from the old version 3.2 technical page. Separates native hybrid/correction/wrapper support and bundled Wasm playback from native browser support; no new WavPack execution is claimed. | 10/8/2026 | 4 |
| WebM | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Replaced the status/web-delivery waffle with a clear list of the allowed codecs (VP8/VP9/AV1, Vorbis/Opus, WebVTT), FFmpeg's refusal of anything else, encode/remux/WebVTT commands and a brief browser-support pointer. Independent review: Quoted FFmpeg error string, VP9/Opus encode, remux, SRT->webvtt mux, mkvmerge .webm output (and its rejection of H.264), and -live/-dash muxer options all verified. Browser paragraph claimed AV1 'depends on whether the browser has a decoder'; Chrome/Firefox/Edge ship software AV1 decoding, Safari decodes AV1 only on hardware with an AV1 decoder (A17 Pro/M3+). Rewritten. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page correctly presents WebM as a constrained Matroska use, clearly distinguishes documented codec subsets from common AV1 implementation behavior, and gives suitably cautious HDR and delivery checks. Earlier pass: 2026-09-07GPT-6 Astra · medium Expanded practical guide with version-scoped recipes, controls, tradeoffs, compatibility, and validation. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted both optional audio maps and added -n to FFmpeg derivative examples. Rechecked current public WebM/W3C subset listings and preserved their distinction from common AV1 implementations. Synthetic VP9/Opus encode/remux/decode and zsh map handling passed. | 10/8/2026 | 4 |
| WebP | current | PassReview basis· claude-opus-5-5 · high · Correction check libwebp 1.6.0 verified current; cwebp/img2webp/gif2webp options and defaults match the v1.6.0 sources. Rewrote all 88 descriptions and corrected facts: -pre bit 4 is sharp YUV, -H is the long help, -near_lossless implies lossless, -z is ignored with -q/-m, -size raises -pass to 6 and -psnr overrides -size, -kmax special values per source. Independent review: No wrong rows in about 35 sampled. Confirmed: -z is ignored if -q/-m appear in either order; -size/-psnr raise -pass 1 to 6; -pre bit 4 is sharp YUV; -alpha_cleanup clears exact; img2webp kmin/kmax INT_MAX-1/INT_MAX; gif2webp kmin/kmax 9/17 lossless and 3/5 lossy (-mixed clears lossless); gif2webp metadata default xmp. Tightened -mt in cwebp and gif2webp ('give it once' was unclear). Coordinator: the appendix now links to the consolidated 'Reading option tables' page. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The technical guidance is otherwise strong, but the cwebp controls table is broken: unescaped vertical bars in the '-preset photo | picture | drawing | icon | text' cell create extra Markdown columns and disconnect the row's description. Earlier pass: 2026-09-07GPT-6 Astra · medium Repaired malformed parameter table and expanded tested lossy/lossless, effort, alpha, and metadata workflow. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Reviews complete libwebp 1.6.0 cwebp/img2webp/gif2webp registrations, hidden controls and defaults. Examples distinguish still lossless hidden-RGB equality, near-lossless changes and animation compositing/timing/loop checks; equal quality numbers, still-only dwebp, metadata and target-device limits are not overstated. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Adds 8-bit lossy 4:2:0 versus lossless ARGB representation, still codec dimensions and frontend/decoder roles. Checks browser feature records and keeps alpha/animation, metadata and still-only dwebp boundaries explicit; preserves complete libwebp inventories. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Corrected a substantive format/implementation limit conflation: VP8 16,383, VP8L size-minus-one fields 16,384, libwebp 1.6.0 encoder API 16,383 in both modes, and extended canvas area/individual-frame boundaries. Kept all parameter references unchanged. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-sonnet-5-5): The page was ~4,000 words of 'Parameter combinations and executed examples' with long ImageMagick cmp verification chains, plus license notes. Rewritten into a practical reference: strengths/weaknesses (8-bit 4:2:0 lossy only, 16,383 px limit, no HDR), browser support from MDN, a minimal `cwebp -q 80 -m 6` command and task table (-sharp_yuv, -lossless -z 9, -near_lossless, -alpha_q, -exact, -size), the metadata/ICC default (`-metadata none`) and Exif-orientation pitfall, animation with gif2webp/img2webp, and CVE-2023-4863. The three bound reference tables (cwebp, img2webp, gif2webp) are kept verbatim at the end. Independent review: 'About 40 % smaller than libpng' was true only against plain libpng; added that Google's same study finds about 20 % (23 %) vs ZopfliPNG-optimised PNG. Removed the unsupported claim that the lossy encoder is 'tuned for PSNR'. Removed 'a few percent more bytes' for -sharp_yuv (unverified). 'Transparent image' row used `-alpha_q 100`, which is the default (ritual); replaced with a 'Smaller alpha' row using -alpha_q 80 and stating alpha is lossless by default. Verified locally (libwebp 1.6.0): default -metadata none drops ICC (webpinfo shows no ICCP) and -metadata icc keeps it; Exif not copied by default; MDN support versions (Chrome 17/23/32, Edge 18, Firefox 65, Safari 14). Coordinator: linked the option appendix to the image encoder parameter reference. | 10/8/2026 | 5 |
| WebVTT | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote around the format (period timestamps, cue settings, ::cue styling), HTML track usage (kind, CORS, text/vtt), FFmpeg conversions and what is lost, and HLS delivery; removed disclaimers and testing narrative. Independent review: False claim: 'FFmpeg's hls muxer writes X-TIMESTAMP-MAP for you'. Neither webvttenc.c nor hlsenc.c at n9.0.2 writes it, and FFmpeg's TS segments start at a 1.4 s offset, so per RFC 8216 §3.5 cues appear early. Rewritten with the RFC rule and -muxdelay 0 (verified it removes the offset). Added RFC 8216 source. ASS->WebVTT keeps bold as well as italics; corrected. Removed duplicated 'hour is optional' sentence. Sample cue file, srt<->vtt and WebM mux tested. Earlier pass: 2026-09-08GPT-6 Astra · not recorded Adds complete WebVTT/HTML examples, correct cue syntax, MIME/CORS and media-codec diagnosis, styling limits and conversion checks; verifies native cue parsing and timed activation in three browser engines, recording the tested Firefox build's need for a WebM media source. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked May20 2026 W3C Candidate Recommendation Draft, external HTML track delivery, MIME/CORS, programmatic cues and packaged-track versus native browser support. No claim of complete styling interoperability. | 10/8/2026 | 5 |
| x264 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrote all 161 standalone rows against x264.c, common/base.c, encoder/encoder.c and ratecontrol.c at b35605ac. Removed the 'Argument: <…>' boilerplate, 'Action; not enabled' and 'as specified by the implementation' text; added short options. Corrected rows: qpmax 81 is the help's 'no cap' value (clamped to 69 for 8-bit, 81 for 10-bit); qp/crf ranges by bit depth; CQP turns off AQ and MB-tree; vbv-init ≤1 is a fraction of bufsize and >1 is kbit; thread-input/min-keyint/sync-lookahead/threads/colormatrix/range auto rules; --stdout/--stdin take an argument (aliases of --muxer/--demuxer); --interlaced; pbratio has no effect with MB-tree on; subme 10 needs trellis 2 and AQ. FFmpeg libx264 table: added a 'maps to and meaning' column, fixed the -forced-idr range (-1..1), and documented -x264opts/-x264-params order and error handling, -passlogfile vs -stats, and the -crf/-qp interaction with -pass 2. Hashes unchanged (local help matches). Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page accurately treats x264 as a rolling source project, pins the current master commit, distinguishes standalone and FFmpeg interfaces, and avoids cargo-cult parameter advice. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks 168 parser names grouped into 161 native rows and 48 FFmpeg 9.0.2 private options, with aliases, defaults, build/preset limits and original explanations. Seven encode/mux recipes decoded completely; the lossless sample reproduced input YUV exactly. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Opened with --version/--fullhelp/-h encoder preambles, a pinned-snapshot narrative, a testing narrative, a 'Validation and reproducibility' checklist and padded examples (--threads 2, -map, BT.709 flags); the original community guidance (aq-mode 3, bframes/ref, MB-tree, open GOP, lossless, MKV muxer caveat) had been stripped. Rewritten as a practical guide: encode/pipe commands, rate-control table (CRF/2-pass/VBV/QP), --level 4.1 ref-capping advice, presets, tunes table from --fullhelp with a psnr/ssim metric-only caution, a verified options table (psy-rd, aq-mode/strength, deblock, mbtree/qcomp, bframes, b-adapt, ref, me, subme, merange, rc-lookahead, trellis, keyint, open-gop, threads), 10-bit/4:2:2/4:4:4 (High 10 hardware caveat; -strict -1 still needed for >8-bit Y4M in FFmpeg 9.0.2), and lossless (--qp 0; --crf 0 is lossless only at 8-bit). Bound parameter-reference sections kept verbatim at the end with shortened intros. Independent review: Verified against r3222 --fullhelp and runs: preset contents (umh/b-adapt 2 start at slower, not slow), tune table, psy-rd/subme/trellis dependencies, --level 4.1 caps refs to 4 at 1080p veryslow (5.1 without), --crf 0 lossless only at 8-bit (High 10 at 10-bit) while --qp 0 gives High 4:4:4 Predictive, FFmpeg 9.0.2 still needs -strict -1 for 10-bit Y4M, x264 Matroska writer has no Cues, artifacts.videolan.org builds exist. Added the ±6 CRF ≈ 2× bitrate rule to the CRF paragraph and changed '10-bit build' to '10-bit output' (current builds are combined 8/10-bit). Added a community starting point (--preset veryslow --tune film --crf 18, + --aq-mode 3 for banding in dark scenes) and noted aq-mode 3's bitrate cost; command tested. Coordinator: removed inline code formatting from a table header row, which failed WCAG colour contrast on the themed table headers. | 10/8/2026 | 7 |
| x265 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrote 274 standalone rows and 11 FFmpeg rows against param.cpp, x265cli.cpp/.h and Encoder::configure at tag 4.3. --selective-sao now reads '0 stored; 4 (all slices) when SAO is on' (stays 0 with --no-sao). Other stored-vs-effective defaults now give both values: min-keyint, lookahead-slices, frame-threads, psy-rdoq (forced to 0 when rdoq-level is 0), mcstf-ref-range (2, or 1 when --preset ultrafast–medium is given), asm (AVX-512 off). Corrected: recon-depth is the output depth; --pmode/--pme are rejected in 4.3 (only the --no- forms parse); --ratetol, --no-dither and --[no-]fast-cbf are rejected; --cu-stats and --analysis-reuse-mode are deprecated; VUI fields are 'not signaled' by default; --dolby-vision-profile also accepts 8.4; --tune includes animation; garbled defaults for opt-qp-pps/opt-ref-list-length-pps fixed. Mislabelled groups fixed. FFmpeg -dolbyvision now correctly says it regenerates RPUs from the input's Dolby Vision metadata. The --fullhelp sha256 was updated to this 4.3 build's stdout hash. Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page correctly identifies x265 4.3 and its canonical repository, explains interface and HDR/container boundaries, and keeps settings conditional and reproducible. Earlier pass: 2026-10-01GPT-6.1 Sol · xhigh Checks 350 source parser names grouped into 274 native rows and 11 FFmpeg-private controls. Corrects help-only fake negations and default spillovers, including ordinary log level and mcstf defaults. Eight encode/mux examples passed strict full decode; lossless output matched source YUV. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Quoted one optional FFmpeg -map value so the question mark remains literal under zsh filename generation. Stream-selection meaning is preserved, and Bash and zsh argument checks passed. Existing encoder or filter behavior was not newly exercised. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Opened with build/--fullhelp/-h preambles, pinned-commit narrative, a testing narrative and validation checklist; examples were padded with --pools 2 --frame-threads 1 and caveats. The old page wrongly claimed lossless needs an extension profile (x265 --lossless output is Main). The original's b-adapt table was wrong (listed modes 1–3; actual 0 none/1 fast/2 trellis, default 2) and its AQ table omitted mode 0. Rewritten with status (4.3, 31 July 2026, actively developed, GitHub upstream), encode/pipe commands, FFmpeg -x265-params and -tag:v hvc1, rate control (CRF/2-pass with slow first pass/VBV), presets, a tunes table including the undocumented-in-help 'animation' tune and grain details, a psnr/ssim metric-only caution, 10-bit/HDR10 command, and verified option tables (psy-rd, psy-rdoq with its rdoq-level dependency and slow+ default 1.0, aq-mode/strength, cutree, hevc-aq, sao/limit-sao/selective-sao, deblock, rd, rdoq-level, tskip, rect/amp, bframes, b-adapt, bframe-bias, ref, me, subme, merange, rc-lookahead, keyint, open-gop) plus threading and lossless. Bound reference sections kept verbatim at the end. Independent review: Tune table: --tune psnr does NOT disable CU-tree in 4.3 (param.cpp sets aqStrength 0, psyRd 0, psyRdoq 0 only; confirmed in the encoded SEI) although presets.rst says it does; fixed psnr/ssim rows to say psy-rd and psy-rdoq off, added 'one frame thread' to zerolatency. Grain tune row now says it costs a lot of bitrate and that moderate grain is usually better served by manual settings; overriding AQ/CU-tree warning kept. Added a tested high-fidelity live-action starting command (--no-sao --deblock -2:-2 --aq-mode 3 --psy-rdoq 2, 10-bit, preset slow, CRF 18) and animation starting points (deblock -1:-1, aq-strength 0.8, +2 B-frames, limit-sao/no-sao), framed as per-source starting points. aq-mode row: reworded the unsourced 'little gain on bright content' to the community caveat that mode 3 can raise bitrate for little visible gain; added --no-strong-intra-smoothing (default on per cli.rst) row; b-adapt default noted as 0 at veryfast–fast (presets.rst). Verified against 4.3 presets.rst/param.cpp/cli.rst/encoder.cpp: psy-rdoq 1.0 at slow+, rdoq-level dependency, rd 3/4 and 5/6 pairs, selective-sao effectively 4 when SAO on, limit-sao, rc-lookahead > keyint not helpful, frame-threads note, slow-firstpass default, crf-max/VBV underrun warning, lossless semantics, 4.3 release date 31 July 2026; added encoder.cpp to sources. | 10/8/2026 | 8 |
| x266 | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Stated plainly that x266 is MulticoreWare's announced (2020), x265-based, consortium-funded VVC encoder with no public source or binaries, removed the 'tracking a future release' checklist and speculation warnings, added a note that the x266.mov domain is unrelated to the encoder, and pointed to VVenC/uvg266/VTM. Independent review: Verified the product page ('built on top of the industry-standard x265 HEVC encoder', consortium invitation) and the 31 Jan 2023 webinar FAQ (license undecided, release at v1.0); no x266 repository in the Multicorewareinc GitHub org. Renamed the defensive heading 'What is actually known' to 'What MulticoreWare has said' (no inbound anchors). Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The page responsibly limits x266 to announced/unreleased status, avoids invented commands and licensing, and points readers to reproducible public VVC alternatives. | 10/8/2026 | 3 |
| XPSNR | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Major correction: FFmpeg's xpsnr filter takes the REFERENCE first (weights come from the first input, per FFmpeg docs and vf_xpsnr.c); the old page told readers to put the distorted file first, which gives different, wrong scores. Verified xpsnr first shipped in FFmpeg 7.1, kept the developers' minimum-of-planes recommendation, described the community 4:1:1 weighted variant neutrally, dropped the unsourced '42 dB visually lossless' threshold, and added the vszip XPSNR filter. Independent review: Reference-first order confirmed in FFmpeg docs ('The first input is considered a (usually not distorted) reference') and by test: swapping inputs changed luma XPSNR from 29.52 to 18.35 dB on the same pair. First release confirmed: xpsnr absent from allfilters.c at n7.0.2, present at n7.1; Fraunhofer repo README says it is now in FFmpeg and unmaintained. 'Report the minimum' matches the FFmpeg docs ('we suggest reporting of the minimum XPSNR average'). Fixed the quoted summary line to FFmpeg 9.0.2's actual format (double spaces) and reworded the pixel-format bullet to what the docs state (bit depth may differ; designed for YCbCr). Verified Av1an's xpsnr-weighted is 4:1:1 (weight_xpsnr in xpsnr.rs). Tested the scale-the-second-input command. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page still instructs users to patch the old Fraunhofer plug-in into FFmpeg 7.0, but XPSNR is now a documented built-in FFmpeg filter; its visually-lossless threshold and weighted-score recommendations are also presented as community anecdotes without adequate support. Earlier pass: 2026-09-07GPT-6 Astra · medium Replaces FFmpeg7 patching with tested built-in filter usage, correct distorted/reference order, alignment checks and component reporting; removes universal transparency and arbitrary weighting advice. | 10/8/2026 | 8 |
| xz | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Condensed to presets (memory table from xz(1)), -T0 trade-offs, tar integration and the fact that xz deletes its input; removed testing narrative. Independent review: xz has been multithreaded by default since XZ Utils 5.6.0 (NEWS; verified: plain 'xz -1' on 10 MB writes 4 blocks, -T1 writes 1). Removed ritual -T0 from commands/XZ_OPT and rewrote the threading paragraph. Memory-limit bullet was wrong: xz has no decompression limit by default; errors only occur when a limit is set. Fixed. Preset memory table matches xz(1) 5.8.4. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The unsourced page makes broad default-platform-support claims and presents a high-effort command without adequately explaining dictionary/memory consequences, stream checks, integrity choices, or how current xz versions interpret the options. Earlier pass: 2026-09-07GPT-6 Astra · medium Adds tested XZ Utils round-trip and documents dictionary/thread memory, automatic adjustments, stream checks and compatibility without platform guarantees. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked XZ5.8.4 auto-thread default, LZMA2/filter/container distinctions, decoder memory and integrity boundaries. Exact byte restoration passed. | 10/8/2026 | 3 |
| Your first FFmpeg encode | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-opus-5-5): Removed the four 'check the tools' commands, the ffprobe field table, -n, -map 0:V:0/'0:a:0?', the per-flag justification list, POSIX/PowerShell note and testing narrative. Page now opens with the idiomatic x264+Opus command, then explains CRF/preset/tune (x264 default CRF 23, ±6 ≈ 2× bitrate), short-clip testing with -ss/-t, FFmpeg's real default stream selection (default-flagged audio wins over channel count; SRT→ASS conversion in MKV), -map usage, copy vs re-encode audio, the libopus 5.1(side) failure and fix, High 10/4:2:2 compatibility and faststart, plus a table of real error messages. Independent review: Default stream selection was described imprecisely (audio 'default flag, otherwise most channels'; subtitles 'if the output can hold it'). Rewrote from n9.0.2 map_auto_*: highest resolution / most channels, a default-flagged stream wins within a file (+5,000,000 score), first subtitle the format's default encoder can handle; MKV converts text subs to ASS and MP4 auto-mapping writes no subtitles (MP4's default subtitle codec is none). Both behaviours reproduced locally. 5.1(side) command lacked -mapping_family 1; added it with the reason (explicit family 1 uses libopus's surround encoder with inter-channel masking; the default -1 writes a family-1 header via the legacy multistream API, and output differs). Added that aformat relabels side→back without downmixing and that -af hits every audio stream (link to per-stream -filter:a:N). Error table: corrected the libopus message to its real text ('… for specified mapping family'); added the MP4 '-c copy' subtitle error 'Could not find tag for codec subrip' next to 'Automatic encoder selection failed'. Unknown encoder and unmatched -map messages verified. -tune description said it only adjusts psy settings; film/animation/grain also change deblocking and AQ (animation also refs/B-frames). Starting command (libx264 slow/CRF 20 + Opus 128k in MKV), CRF 23 default, ±6 ≈ 2× bitrate, -ss before -i accuracy when re-encoding, High 10/4:2:2 pixel-format caveat and faststart advice are sound and kept. Earlier pass: 2026-08-25GPT-5.6 Sol · xhigh The beginner workflow is technically sound: it checks local encoder availability, maps streams deliberately, explains x264 CRF and preset roles without overgeneralizing, and includes decode, probe, playback, and HDR caveats. Earlier pass: 2026-09-08GPT-6 Astra · not recorded Expands source inventory, scoped SDR sampling, explicit stream selection, full-file encoding and acceptance checks; exercises the exact FFmpeg 9.0.1 commands, frame counts, Opus output and overwrite refusal on synthetic media without claiming universal quality or compatibility. | 10/8/2026 | 7 |
| YUView | current | PassReview basis· claude-opus-5-5 · not recorded · Correction check Added a ProjectCard showing the upstream repository, latest release, last commit and licence from the committed forge snapshot (src/data/forge/snapshot.json, fetched 2026-10-08). Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Added tagged-release versus development context, raw-format requirements, filename-inference limits and a packed-size example for 10-bit planar 4:2:0. Distinguished ordinary decoding from decoder-specific analysis overlays. The bounded FFmpeg export passed on 12 synthetic frames at 128×72, producing 331776 bytes; the GUI and instrumented decoders were not run. Earlier pass: 2026-10-08T23:10:44.797Zclaude-opus-5-5 · high Rewrite (claude-opus-5-5): Removed raw-format hedging and the 'GUI was not run' note; rewrote as what YUView does, install (v2.14 still the latest release, Flathub, Ubuntu apt, macOS quarantine fix from the README), raw YUV naming that YUView actually parses (size_fps_pixfmt per its wiki), a tested raw export command, comparison workflow, and the analysis-decoder table (libde265/HM for HEVC, dav1d for AV1, experimental VTM for VVC) with a tested Annex B extraction command. Independent review: Ubuntu claim corrected with Launchpad data: 22.04 ships 2.9 and 24.04 ships 2.13 (2.14 only from 25.04). Simplified Annex B extraction to `ffmpeg -i input.mkv -c:v copy out.hevc` (FFmpeg inserts hevc_mp4toannexb automatically; start code checked); file-name parsing confirmed against the YUView wiki. | 10/8/2026 | 6 |
| ZIP | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote with zip/unzip commands, a methods table (Deflate, bzip2, zstd 93 etc.), ZipCrypto vs AES (unzip cannot read AES zips - confirmed) and ZIP64/filename pitfalls. Independent review: Methods table implied Info-ZIP unzip cannot read non-Deflate methods; tested unzip 6.0 reads Deflate64 and bzip2, skips LZMA and PPMd. Table corrected. 'zip adds and updates changed files' was inaccurate (plain zip adds/replaces; -u updates). 'File names are never encrypted' reworded. AES zip rejected by unzip ('need PK compat. v5.1') confirmed. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page materially confuses Deflate64 method 9 with ZIP64, claiming most compressors use Deflate64 to bypass 4 GiB limits; ZIP64 extensions, not a compression method, provide the larger size and entry fields. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrects Deflate64 versus ZIP64, scopes interoperability by features, and adds tested DEFLATE ZIP creation/test/restore with CRC and encryption caveats. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked zip format capabilities, reference implementations, practical options and resource/integrity boundaries against primary sources. A 37,888-byte mixed binary/text fixture round-tripped exactly with the documented frontend. | 10/8/2026 | 3 |
| zpaq | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote around what zpaq is for (journaling, deduplicated incremental backup, restore any version) with tested add/list/extract/-until/-key/-test commands; built zpaq 7.15 from source for testing. Independent review: Built zpaq 7.15 and verified a/l -all/x -until/-to/-test/-key/-repack, ???? multipart naming and option descriptions against its usage text. Only prettier reformatting. Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh This unsourced stub conflates the ZPAQ format, its journaling archive application, and compression methods, and provides no usable command, version, integrity, restore, or resource guidance. Earlier pass: 2026-09-07GPT-6 Astra · medium Separates ZPAQ format and journaling application, identifies upstream 7.15, and adds a tested incremental-archive creation/list/restore workflow and resource/version-retention guidance. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked zpaq format capabilities, reference implementations, practical options and resource/integrity boundaries against primary sources. Binary unavailable; command shapes are source-checked, with no executed round-trip claim. Earlier pass: 2026-10-04T14:41:35.275Zgpt-6.1-sol · ultra Completed the source-checked fixture with actual extraction and comparison at restored/input.bin, preserving the explicit no-local-execution label and new archive/directory requirement. Clarified stored checksum versus provenance/authentication. | 10/8/2026 | 2 |
| zstd | current | PassReview basis· claude-opus-5-5 · high · Correction check Rewrite (claude-sonnet-5-5): Rewrote with real command examples (levels, -T0, --long, --ultra, --adapt, tar integration) and corrected the old page's .zstd extension (it is .zst) and HTTP guidance (8 MiB window per RFC 9659, so --long/--ultra output is not valid for Content-Encoding: zstd). Independent review: Verified window sizes on a 300 MB input: levels <=19 use 8 MiB, --ultra -20 32 MiB, -22 and --long 128 MiB; --long=30 output is refused by a default decoder and decodes with --long=30. RFC 9659 8 MiB guidance is correct. Browser support was stale ('Safari partial'): caniuse data shows Chrome/Edge 123, Firefox 126, Safari 26.3 (iOS; macOS 26.3+) advertise zstd. Updated and sourced. 'Linux kernel image ... use it' overstated; now 'the kernel, btrfs, SquashFS and OpenZFS support it'. 7-Zip .zst unpack-only since 24.01 confirmed in history.txt and locally (7z a -tzstd: E_NOTIMPL). Earlier fail: 2026-08-25GPT-5.6 Sol · xhigh The page is materially stale: it describes browser zstd content encoding as only an experimental Chromium 118 possibility, and its command text uses -ultra instead of the documented --ultra option. Earlier pass: 2026-09-07GPT-6 Astra · medium Corrects --ultra, adds tested round-trip and dictionary/window tradeoffs, and replaces obsolete browser speculation with registered negotiated HTTP coding and its window limit. Earlier pass: 2026-10-04gpt-6.1-sol · ultra Checked zstd format capabilities, reference implementations, practical options and resource/integrity boundaries against primary sources. A 37,888-byte mixed binary/text fixture round-tripped exactly with the documented frontend. | 10/8/2026 | 5 |
Maintained refreshes
These migrated pages have been deliberately promoted out of generated migration ownership. Their reviewed edits survive a complete source-corpus regeneration.
| Page | Status | Last verified | Promotion reason |
|---|---|---|---|
| 7z | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| AAC | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| ALAC | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| AMF | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| Antialiasing | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| aom-av1-lavish | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| aom-psy101 | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| aomenc | current | 10/8/2026 | Corrected the stale archived status and refreshed the page around the libaom v3.14.1 reference encoder and current build interface. |
| Aurora1 | current | 10/8/2026 | Replace a vendor stub with source-checked evaluation guidance and explicit SDK verification limits |
| Autocompressor | current | 10/8/2026 | Replace migrated marketing prose with dated hosted controls, a target-size workflow and explicit verification limits. |
| AV1 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| Av1an | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| av1an-command-gen | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| AVC / H.264 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| Aviator | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| AVIF | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| AVM reference software | current | 10/8/2026 | Replace pre-standardization rumors with a versioned AV2 v1.0.0 reference-software guide. |
| AVS3 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| Basics of Filtering | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| brotli | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| Butteraugli | current | 10/8/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| bzip2 | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| Color Formats | current | 10/8/2026 | Correct endianness, pixel-format names, component precision, and unsupported subsampling advice; approved by the site owner for publication on 4 October 2026. |
| Color Primaries | current | 10/8/2026 | Correct migrated color-signaling errors using dated primary specifications and source definitions; independent human review remains pending. |
| Color Range | current | 10/8/2026 | Expand range interpretation, metadata-versus-pixel conversion and tested SDR diagnostics |
| Colorimetry: inspect before encoding | current | 10/8/2026 | Correct migrated color-signaling errors using dated primary specifications and source definitions; independent human review remains pending. |
| CVVDP | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| dav1d | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| Deband | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| Decombing | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| Dehalo | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| Deinterlace | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| Denoise | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| Discord | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| Dolby audio codecs and Atmos | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| dovi_tool | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| DSV | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| DSV2 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| eac3to | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| ECM | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| Encoding basics | current | 10/8/2026 | Editorial rewrite: verify technical claims against current primary sources, remove unsupported claims and padded commands, and restore practical community guidance. |
| Eve-AV1 | current | 10/8/2026 | Replace unversioned migrated claims with a source-pinned October 2026 encoder parameter/default reference and explicit verification limits. |
| Eve-VP9 | current | 10/8/2026 | Document public commercial encoder scope, unpublished parameter/default boundaries and reproducible VP9 output checks. |
| exhale | current | 10/8/2026 | Add a complete version-pinned xHE-AAC CLI reference with executed examples and explicit separation from AAC-LC tools. |
| FAQ | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| FFMetrics | current | 10/8/2026 | Reviewed against the v1.7.0 release, current metric list, and documented 24-input comparison workflow. |
| FFmpeg | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| FFV1 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| FLAC | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| GIF | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| Graining | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| gzip | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| hdr10plus_tool | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| HEIC | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| HEVC / H.265 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| High dynamic range | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| HM | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| Intro | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| Inverse Telecine | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| JM | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| JPEG | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| JPEG 2000 | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| JPEG XL | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| Kvazaar | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| Lossless compression | current | 10/8/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Lossy compression | current | 10/8/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Matrix Coefficients | current | 10/8/2026 | Correct migrated color-signaling errors using dated primary specifications and source definitions; independent human review remains pending. |
| MediaCodec | current | 10/8/2026 | Replace unversioned migrated claims with a source-pinned October 2026 encoder parameter/default reference and explicit verification limits. |
| MKVToolNix | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| MP3 | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| MP4Box | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| NMKODER | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| NVENC | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| Opus | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| PNG | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| ProRes | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| PSNR | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| Psychovisual encoding | current | 10/8/2026 | Editorial rewrite: verify technical claims against current primary sources, remove unsupported claims and padded commands, and restore practical community guidance. |
| qaac | current | 10/8/2026 | Add a version-pinned complete AAC frontend parameter reference with explicit separation from proprietary runtime capabilities. |
| QOI | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| QSV | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| rAV1ator | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| rAV1ator CLI | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| rav1e | current | 10/8/2026 | Replaced obsolete 0.7.1 guidance with a dense review of the current stable v0.8.1 release, interfaces, and reproducibility requirements. |
| Resources | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| Speex | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| SRT | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| SSIM | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| SSIMULACRA2 | current | 10/8/2026 | Editorial rewrite: verify technical claims against current primary sources, remove unsupported claims and padded commands, and restore practical community guidance. |
| Stabilizing | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| SubStation Alpha and ASS | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| SVT-AV1 | current | 10/8/2026 | Rewritten from the migrated page to document the AOMedia mainline v4.2.0 interface, limits, and reproducible usage. |
| SVT-AV1-PSY | current | 10/8/2026 | Historical entry retained for lineage and reproducibility while clearly marking the original project as archived and directing new work to maintained successors. |
| SVT-HEVC | current | 10/8/2026 | Rewritten as an archived page after confirming the upstream project is discontinued and its latest release is v1.5.1 from 2021. |
| SVT-VP9 | current | 10/8/2026 | Rewritten as an experimental implementation page after checking SVT-VP9 v0.3.1 and its upstream user guide. |
| tar | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| Terminology | current | 10/8/2026 | Editorial rewrite: verify technical claims against current primary sources, remove unsupported claims and padded commands, and restore practical community guidance. |
| Theora | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| Transfer Characteristics | current | 10/8/2026 | Correct migrated color-signaling errors using dated primary specifications and source definitions; independent human review remains pending. |
| uavs3e | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| UT Video | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| uvg266 | current | 10/8/2026 | Rewritten from the migrated page after reviewing the v0.8.1 release and current upstream README; documents that the latest tagged release is older than the active development branch. |
| VapourSynth | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| VC-1 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| Video artifacts | current | 10/8/2026 | Editorial rewrite: verify technical claims against current primary sources, remove unsupported claims and padded commands, and restore practical community guidance. |
| Video Players | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| VideoToolbox | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| VMAF | current | 10/8/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Vorbis | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| VP8 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| VP9 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| vpxenc / libvpx | current | 10/8/2026 | Rewritten as the shared VP8/VP9 implementation page after reviewing libvpx v1.15.2 and removing unverified 2021-era universal recommendations. |
| VTM | current | 10/8/2026 | Replace failed migrated guidance with sourced September 2026 explanations, usable commands, and explicit validation limits. |
| VVC / H.266 | current | 10/8/2026 | Expand October 2026 codec capabilities, profile and chroma boundaries, implementation/decode coverage, and dated browser support guidance. |
| VVenC | current | 10/8/2026 | Rewritten from the migrated page after reviewing VVenC v1.14.0, its usage guide, and current FFmpeg integration documentation. |
| WavPack | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| WebP | current | 10/8/2026 | Extend October 2026 sourced codec capability, implementation and browser support guidance while preserving parameter references. |
| WebVTT | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| x264 | current | 10/8/2026 | Rewritten as an active rolling x264 page after reviewing the VideoLAN source repository and current master commit; unsourced fastest/best claims were removed. |
| x265 | current | 10/8/2026 | Rewritten from the migrated page after reviewing the authoritative x265 4.3 repository, release, and CLI documentation. |
| x266 | current | 10/8/2026 | Rewritten to correct the lifecycle status after checking MulticoreWare's public FAQ and current product material; no public source or release was verified. |
| XPSNR | current | 10/8/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| xz | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| YUView | current | 10/8/2026 | Replace unsupported migrated guidance with dated primary evidence and practical usage; approved by the site owner for publication on 4 October 2026. |
| ZIP | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| zpaq | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| zstd | current | 10/8/2026 | Expand capabilities and useful versioned parameters with explicit execution and renderer limits |
| Better late than never: SVT-AV1 v2.2.x Deep Dive | current | 10/4/2026 | Qualify historical version and harmonic-pooling limits without changing preserved charts or author attribution; independent human review remains pending. |
| Codec Wiki: One Year Later | current | 10/4/2026 | Add historical project and analytics context while preserving the author report and original screenshots; approved by the site owner for publication on 4 October 2026. |
| Embedding the Un-Embeddable | current | 10/4/2026 | Qualify dated platform observations and distinguish video metadata from actual playback support; approved by the site owner for publication on 4 October 2026. |
| Encoding Animation with SVT-AV1: A Deep Dive | current | 10/4/2026 | Qualify historical benchmark conclusions and preserve the attributed experiment and charts; approved by the site owner for publication on 4 October 2026. |
| AV1 Encoding for Dummies | current | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| AV1 for Dummies | current | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Contributing after the Codec Wiki migration | current | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Deep Dive into SVT-AV1's Evolution (Part 1): Presets Analysis from v2.0 to v3.0 | current | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Deep Dive into SVT-AV1's Evolution (Part 2): Encoder Parameters Revisited | current | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Fast Computation of SSIMULACRA2 on GPUs: A Performance Evaluation | current | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Historical Codec Wiki privacy policy | archived | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Historical Codec Wiki terms of use | archived | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Observing SVT-AV1 v2.1.0's improvements: A New Deep Dive | current | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |
| Reducing Image Load Online | current | 9/7/2026 | Correct the August 2026 review findings with dated primary evidence while preserving attribution and historical experimental material. |