Repository navigation
Description
Goal
coreav ffprobe prints what the pinned ffprobe prints: bar B3a of docs/design/compat-cli.md ("The bars, pre-registered"), every key equal on at least 98% of the corpus files it appears in and on 100% of the generated fixtures. Until a round on new code meets it, the command stays labelled experimental.
Current state
-
feat(ffprobe): summarize bounded hev1 inputs #430 merged native input summaries for qualified video-only HEVC
hev1MP4 files. Admission requires complete linked parameter sets, one temporal layer and one selected slice in every sample, with strict bounded prefix validation. Generated full-summary, structured-output and fallback checks match FFmpeg 9.0.2; unsupported layouts retain atomic refusal or original-command fallback. Existinghvc1and structured behavior are unchanged. This does not establish a new B3a or command-coverage verdict. -
feat(ffprobe): read JSON probes from stdin #412 merged complete native stdin JSON for qualified front-loaded, nonfragmented ISO movies smaller than 1 MiB, using
pipe:orpipe:0. Logical pipe filenames are retained while unknown format size and total-file bitrate are omitted. Unsupported or cap-exhausted prefixes relay all consumed bytes and live stdin to fallback without waiting for producer EOF. Generated controls cover startup before EOF, complete replay, exit status and cleanup. This does not expand other container layouts or establish a new coverage or B3a verdict. -
feat(ffprobe): qualify unspecified H.264 aspects #394 merged bounded native ISO file input summaries. Complete codec-configuration and sample-reference authority permits explicit unknown aspect values without claiming decoded-picture integrity. Generated checks pass, and the complete 926-input development diagnostic moves from 233 to 254 exact summaries with no lost exact result, mismatch or incomplete input. Unsupported inputs retain atomic refusal and complete pinned fallback. This is diagnostic evidence, not a new B3a or published command-coverage verdict.
-
fix(ffprobe): correct ISO format durations #371 merged the ISO format-timing correction. The ffprobe output uses eligible ancillary bounds and the one-second cutoff while preserving source stream timing, raw probe values and stream-copy planning. Generated controls and a non-holdout development reproduction match pinned FFmpeg 9.0.2 on the selected format fields; copied-output status and hashes remain unchanged. This is limited diagnostic evidence, not a new B3a or GitHub-shape verdict.
-
fix(ffprobe): correct ancillary stream rates #375 merged the ISO ancillary stream-rate correction: data and subtitle estimates use their encoded media span rather than presentation edits. Flat rates match pinned FFmpeg 9.0.2 on 43 generated/development controls, with other native metadata, raw probes and nine copied-output hashes unchanged. Selected command stdout is exact for 99 of 129 controls; 30 retain an existing timecode stream-group listing gap and are not counted as passes. This is a limited diagnostic, not a new B3a corpus verdict.
-
feat(ffprobe): probe standalone audio natively #332 merged bounded native standalone PCM WAVE and FLAC format, stream and text-tag probing. Pinned FFmpeg 9.0.2 comparisons pass 45 generated fixtures, 82 byte-identical default/JSON comparisons, six strict-safe refusals and two exact divergence assertions. Unsupported inputs and packet/frame inspection retain fallback. Extensible PCM uses its specified storage width where the oracle disagrees; the public divergence reference records that exception. No new corpus-agreement or speed verdict is claimed.
-
docs(bench): record B4 development calibration #305 records the passing round 10 development-only B4 calibration: both registered controls pass, all 4,992 calls complete, and no failures or fallback records occur. Complete public-safe receipts reproduce the report exactly. No product or held-out measurement ran, and no product speed verdict is established. Historical records and bars remain unchanged.
-
feat(bench): add counterbalanced B4 driver #283 merged the registered
b4-counterbalanced-v2driver. It freezes development and product rosters with held-out disjointness, schedules counterbalanced calibration, captures resources outside the shared timer and validates complete global receipts before reporting. Calibration-only mode never opens product inputs. Historical records, product bars and R4 are unchanged. No new product speed verdict is claimed. -
docs(bench): register B4 calibration #257 merged the future round 10 B4 calibration plan: counterbalanced crossovers, repeated-round and file resampling, fixed conditioning and outside-timer telemetry. Its synthetic coverage and power evidence is committed. Product bars, predicates, commands, estimators and whole-process timing stay unchanged; round 9 is not re-scored. The separate driver merged in feat(bench): add counterbalanced B4 driver #283; measurement still requires its source and host gates. No new product speed verdict is claimed.
-
perf(ffprobe): reuse corpus audit evidence #217 merged bounded reuse of byte evidence, decoded packet rows and baseline audit coverage. Reuse validates media and reader byte identities and rechecks attacked occurrences without changing the frozen exclusions or judgments. Development replay was byte-identical; this is not a new corpus-agreement or speed verdict.
-
docs(ffprobe): bound corpus key exclusions #150 and feat(ffprobe): audit corpus key exclusions #157 merged the round-9 exclusion plan and exact-occurrence comparator/audit tools. Seven byte-validated adapters cover three MPEG-TS cases, three Matroska/WebM cases and the single ISO empty-edit padding key. Records and scopes freeze at the plan amendment; raw agreement remains beside adjusted rates and separately attributed exclusions. Thresholds, inputs and historical results are unchanged. The judged run has not started and requires the full main-CI and host gates.
-
fix(ffprobe): list ISO audio priming and padding #156 and test(ffprobe): read empty-edit padding from file bytes #160 merged ISO audio priming/padding and byte-only empty-edit evidence. The decoded AAC frame span comes from its configuration; skip and padding use the output rate. Generated empty-edit exceptions require their manifest identity and exact byte-derived output pair. The corpus helper supplies per-packet evidence, including a tightly bounded proven-absence case; the round-9 comparator consumes it through feat(ffprobe): audit corpus key exclusions #157. fix(ffprobe): list no cover without a meta handler #167 also removed attached-picture streams from metadata boxes without their mandatory handler. Generated fixture gates and CI pass; no new corpus-agreement or speed-bar verdict is claimed.
-
docs(ffprobe): register the next validation arm #115 merged the next validation plan without changing the bars. Round 8 remains a B3a miss: keys below 98% changed from 28 to 25 of 106 for ISO base media, 25 to 22 of 113 for Matroska/WebM, and 44 to 39 of 177 for MPEG-TS. Its B4 run is not a clean protocol verdict; it applied a forced MPEG-TS command to non-TS files and resampled command/input pairs rather than whole files. It is not rerun or re-scored. The held-out MPEG-TS files inspected during this work are no longer independent validation.
-
fix(ffprobe): follow pinned stream metadata #116 merged with the stream-vocabulary, timing, QuickTime vendor-tag and VP8-in-MP4 pixel-format fixes, preserving main's VP8 decoder and a shared configuration-record reader. AV1 strings follow the binding's colour/chroma rules where the oracle disagrees. One generated fixture has a byte-validated exact-pair exception that fails when the oracle starts agreeing; no corpus input is excluded. The correction batch fixes LATM and Dolby field ordering, zero-duration text tails and PES-identifier provenance. The instruction guard passes without changing its baseline or tolerance; CI and independent review cleared the merged head. This is not a new corpus or speed verdict.
-
fix(ffprobe): enforce the registered timing method #117 merged the separate measurement-tools change. Its registered file-cluster intervals use 10,000 resamples, forced-format eligibility is decided before execution, failed calls have no ratio, and refusal records are distinguished from executed fallbacks. B4 uses an event-driven process-exit deadline rather than timeout polling; real-child regressions cover timing and cleanup. The upstream code prerequisite feat(mp4): copy FLAC, ALAC, AC-3 and E-AC-3 audio #109 is on main; round 9 still requires the full main-CI and host gates. No new agreement or speed result is claimed; results will be published separately.
-
fix(isobmff): take FLAC and ALAC rates from their config #113 fixes FLAC and ALAC sample rates in MP4 from codec configuration, including 88.2, 96 and 192 kHz, and FLAC bit depth/sample format. Generated fixtures and malformed-configuration regression tests cover the change; 16/24-bit FLAC and ALAC at 44.1, 48, 88.2, 96 and 192 kHz match both pinned ffprobe versions on those fields. This does not rerun the corpus agreement or speed bars.
Round 7 (docs/measurements/compat-ffprobe.md, #59):
- B3a on the fixtures is met in all three containers (ISO base media, Matroska and WebM, MPEG-TS).
- B3a on the corpus is missed: 9 of 101 ISO base media keys and 35 of 157 MPEG-TS keys agree on fewer than 98% of files (keys outside the tags, in at least five files). Matroska and WebM: 0 of 101. Round 6 had 34, 35 and 2.
- fix(ffprobe): agree with ffprobe on more corpus keys #59 closed the ISO base media and Matroska root causes behind most keys: codec fields of entries without a configuration record, display matrices with the movie matrix, chapters and chapter pictures, covers, durations with a leading empty edit, and laced Matroska audio. ISO base media still misses on
streams.r_frame_rate(95.52%),packets.duration(97.24%) and the chapter keys on one file. The held-out ISO base media and Matroska files are not an independent check of this work; only the held-out MPEG-TS files are unseen. - MPEG-TS did not move in round 7 (development 6 of 157, held out 35 of 151), fix(compat): MPEG-TS lead bytes of every audio syntax #39 included.
- B1 (refusals) is met; B2 misses on refusals by design and on fixtures CoreAV rejects, not on wrong values.
- B4 speed was met in round 3: median ratio 0.244 against
ffprobe(95% interval 0.240 to 0.251); not re-run since. tools/compat-corpus/record.py --check(test(compat): tie the ffprobe record to its JSON #35) ties the per-key tables and the figures ofcompat-ffprobe.mdto each round's JSON in CI.
Checklist
-
Qualify bounded native ISO file input summaries with complete configuration and sample-reference authority in feat(ffprobe): qualify unspecified H.264 aspects #394.
-
Qualify complete structured stdin JSON with logical pipe facts, atomic refusal and complete fallback replay in feat(ffprobe): read JSON probes from stdin #412.
-
Extend no-section input summaries to other layouts only after qualifying their transport timing and codec authority.
-
Qualify bounded video-only HEVC
hev1input summaries in feat(ffprobe): summarize bounded hev1 inputs #430 without weakening existing admission or structured output. -
Qualify complete parameter-set authority for existing
hvc1summaries independently of readable metadata prefixes. -
Evaluate HEVC picture-sequence ordering across samples before extending summary conformance guarantees.
-
Check
hvcCtemporal-layer hints against the SPS before extending configuration-record consistency guarantees. -
Add explicit emitted-video start and duration assertions to the delayed-video format-span isolation regression.
-
Correct non-fragmented ancillary stream-rate estimates independently of edits in fix(ffprobe): correct ancillary stream rates #375.
-
Complete ISO timecode stream-group listings for selected stream JSON/default/CSV requests without changing their already matching flat fields.
-
Evaluate native MP3 and Ogg probing with generated-file parity evidence before deciding their scope.
-
Evaluate richer standalone audio metadata, including nonempty ID3 and currently unsupported WAVE and FLAC metadata.
-
Evaluate bounded FLAC frame-derived timing and packet/frame inspection without weakening the header-only fallback boundary.
-
Register the future B4 calibration design and synthetic validation in docs(bench): register B4 calibration #257.
-
Implement and verify the registered round 10 B4 driver in feat(bench): add counterbalanced B4 driver #283 before any new measurement.
-
Record the passing registered development-only B4 calibration and replayable receipts in docs(bench): record B4 development calibration #305, without a product speed verdict.
-
Normalize malformed B4 report JSON to the documented invalid-input exit status without changing rejection or measurement predicates.
-
Add remote completion notification so collection and cleanup begin promptly without a continuously running controller.
-
Establish calibrated B4 capacity before the next judged validation; a rejected calibration does not establish a speed result.
-
Reconcile the frozen reader digest registration without changing reader behavior or exclusion coverage.
-
Merge the registered B4 measurement tools in fix(ffprobe): enforce the registered timing method #117.
-
Merge the frozen corpus-exclusion plan in docs(ffprobe): bound corpus key exclusions #150 and comparator/audit tools in feat(ffprobe): audit corpus key exclusions #157.
-
Reuse independent byte evidence between comparison and mutation audit without widening exclusion coverage.
-
Add bounded reuse of parsed packet rows and baseline audit coverage while revalidating each attacked occurrence and packet identity.
-
Validate under the merged round 9 plan and tools after the full main-CI and host gates; publish the results separately.
-
Replace R4's separate polling timeout wait before its next measurement; the B4 tools intentionally leave R4 unchanged.
-
Check implicit mono HE-AAC/PS and LATM HE-AAC against the pinned oracle; the explicit-extension fixes do not establish their closure.
-
Close the remaining bit-rate, gapless, timecode-group and proprietary highlight-atom differences with specification and byte evidence.
-
Close the ISO base media keys below 98%.
-
Close the MPEG-TS keys below 98%, checked on the held-out transport streams.
-
(fix(ffprobe): agree with ffprobe on more corpus keys #59) Close the two Matroska and WebM keys.
-
(fix(compat): MPEG-TS lead bytes of every audio syntax #39) A frame cut at exactly seven bytes after a stream's lead bytes gets no presentation or decoding time: the lead-bytes packet seeds no next time (10 keys on two crafted AC-3 and E-AC-3 streams, fix(compat): MPEG-TS packet listings closer to ffprobe #17).
-
(fix(compat): MPEG-TS lead bytes of every audio syntax #39) List the bytes before a first frame header for AAC, DTS and MPEG audio as well, where the PES packet cuts the header too short to read: fix(compat): MPEG-TS packet listings closer to ffprobe #17 did it for AC-3 and E-AC-3 only, so those streams still lose one packet and shift the rest.
-
Two held-out transport streams disagree on every packet key (12 of 14 files agree); they hold the largest part of what MPEG-TS has left. One is a multi-program stream whose listing parts from
ffprobe's within its first 30 packets, and which also differs on stream parameters (a video stream's frame rates, an audio stream's sample rate, channels and profile, durations), so its streams are read differently before any packet is placed. The other agrees until packet 17,433, where the listings part, and also differs on its first stream'sclosed_captions. -
ISO base media
streams.r_frame_rate: modelffprobe's base-rate guess for tracks without a regular timing grid (for example H.264 without VUI timing reported at 120/1, HEVC at 30000/1001 instead of 30/1). -
ISO base media
packets.duration: the remaining tail, among them the first packet of some AAC tracks and the last packet of some raw PCM tracks. -
ISO base media chapter keys: the one corpus file whose chapters still differ.
-
Implicit-SBR HE-AAC in MP4 (
sample_rate),extradata_sizeof SVQ1 and QDM2, BMP covers (type 27), and the packet durations of encrypted WebM Opus. -
Keep the chroma format and bit depth fields of an H.264 configuration for profile 244 in the sanitizer and the Matroska-to-MP4 writer; today they follow the 100/110/122/144 list of ISO/IEC 14496-15 and drop them.
-
Run a new round on new code, with B4 measured again.
-
Remove the experimental label once a round meets B3a.
-
Reassess ISO crop keys and
format.nb_stream_groupsafter the merged metadata changes, using specification and byte evidence. -
Reassess ISO SAR/DAR and
mime_codec_stringformp4v.20after the merged metadata changes. -
Audit the ISO
streams.tags.vendor_idomission against the specification and pinned oracle. -
Check
streams.bit_rateon fragmented ISO movies; keep this separate from the non-fragmented audio-padding fix.
Not in scope
- No-section input-summary layouts beyond the qualified native ISO file subset.
coreav ffmpeg:coreav ffmpeg: common-c copyinvocations #25.
Activity
docs(compat): B3a round 6 on the corpus and the fixtures (#21)
test(compat): refresh Matroska and MPEG-TS baselines for ffprobe 9
test(compat): refresh the ISOBMFF baseline for ffprobe 9
capy-ai commentedon Nov 1, 2026
Current state update: #169 merged as a5bca65. KD-FF9-ANNEX-D-DOWNMIX now records the Matroska/WebM AC-3 Annex D downmix disagreement with byte-derived evidence across every frame, including fail-closed traversal and size limits. This changes no CLI output and excludes no corpus rows yet; the comparator must consume the evidence before the record can cover a comparison.