| Filename | Latest commit message | Latest commit date |
|---|---|---|
research/catalogue-prior/decisive-rule.json is the rule that judges a run on a learned prior tuned on a catalogue (H1c) or on an episodic series (H1b), committed before any such run. Its SHA-256 is 8e265cfa2a3cb3de8c02a2c79a65bba59b52f2c134f4baf39d8bf163c0c38bc2. It keeps the record's rule and answers the three faults §2.1 lists: K1 is diagnostic, the bar for a video claim is plain per-title AV1, and K4 kills H1c for an arm whose catalogue-tuned set beats the same codec's generic set by less than 5 % BD-rate on PSNR on the median held-out title, before the set's bytes are charged. Its `base` field names what is new in this file and not carried over: - GO judges H1b on the openings arm, against per-episode encodes, where the record's rule named no comparator for H1b and only K3 decided it. - Each clause names the arms it judges. GO judges wai.neural.int_hyper's catalogue-tuned set against AV1 all-intra, the catalogue-tuned set of any other arm with inter prediction whose P-frames are not integer block motion compensation, named in a run's record before its first reading, against plain per-title AV1, and the openings. K4 judges each catalogue-tuned set against its generic set. wai.video.int_hyper's verdict is a kill or INCONCLUSIVE. - One AV1 bar: 8-bit at preset 4, coded as §3.5 codes base8 in all else, and all-intra at preset 4. Preset 8 and a ×4 10-bit pipeline are reported. - A kill wins over GO for an arm and claim, and K1 with GO is INCONCLUSIVE. K1 reads only a rate proxy measured in the same run on the same titles, so a run that measures none has no K1, and run.json's K1 carries to no run here. - GO's three quarters are of all the arm's titles, rounded up. A title INCONCLUSIVE on VMAF stays among them and does not count as meeting the bar. GO reads the cubic VMAF figure. tools/wai_catalogue_prior_verify.py holds each rule file's SHA-256 itself (RULE_SHA256), treats every file of the record named *-rule*.json as a rule file, and checks it against that digest and the one §2.2 quotes in a <!--cat-sha256:NAME--> marker. Its --self-test requires the record to pass and each mutated copy to fail with its finding: a byte of the rule, a quoted digit, the rule and its quoted digest changed together, an unlisted rule file, a removed one, a removed, stray or misplaced marker, another format, and text that is not JSON. The conformance step runs it before the record. Co-Authored-By: Claude Opus 5.5 <[email protected]> |
||
| .. | ||
| corpora | ||
| verifiers | ||
| baseline.json | ||
| README.md | ||
What wai-conformance.yml runs, one file a step
wai-conformance.yml keeps no hand list of its verifier steps, corpora or kernel tests. Its
lists are derived, when it runs, from the files here and from tags in the test files, by
scripts/wai-conformance-lists.py, whose
docstring states every key; its gate reads the steps' paths: through
scripts/pr-touches-paths.py. A change that adds a verifier, a corpus or a kernel test adds
a file, or a line to a file of its own, and edits no line of the workflow, unless a step's
paths: names a file under none of the trigger's paths (check names it). Its features and
[[test]] entries in wai/wai-rs/Cargo.toml stay shared
(wai/CONTRIBUTING.md).
verifiers/<name>.sh: one step of theindependent-verifiersjob. Its body is the step's shell, run inwai/(orworking-directory:) withbash -eo pipefail, as a workflow step runs. Its header, the leading# key: valuelines, givesstep:(its name),order:(steps run in ascending order, a tie by file name),features:(each list compiles alone infeatures-alone, and every feature joinsWAI_FEATURES), and where it needs themalone:,build: pq,release-features:,bins:,pq-bins:,paths:(the extension texts it reads, which the workflow's gate then runs on) andworking-directory:. Prose comments follow the header, after a#line.corpora/<name>.txt: one cargo-class corpus, verified by its own tool on every leg ofcargo-classes(corpus:,order:,bin:,dir:).- A test file of
wai/wai-rs/tests/whose first line is// wai-conformance: int-kernelsruns in the fast integer kernels' steps of both profiles (int-kernels-release: the release step only;int-kernels-amd: the AVX-512 host too). The steps' features are the tagged tests'required-features. baseline.json: the lists the workflow kept by hand before they were derived. The repository checks require every derived list to hold every entry of it, and every step it names to run.
To see what the workflow runs: python3 scripts/wai-conformance-lists.py dry-run (of another
commit: --ref <commit>; two of them diff line by line). To run one step as CI does:
python3 scripts/wai-conformance-lists.py run --only <name>. To check the files:
python3 scripts/wai-conformance-lists.py check.