Replaying an accepted visual reference batch
corpus-tool compare can reuse an independently accepted
set of reference PDFs, PNG pages and receipts on a machine without the
original reference applications. This is useful when the original batch
lives in an archive and the current native renderer runs in an isolated
CI container.
The operator must first establish the acceptance and provenance of that batch. Creating a manifest and calculating its hash does not establish that its reference PDFs are correct. Preserve the original producer receipt, independent font/page audit, input inventory, exclusions and their trusted hashes.
Pass both --frozen-references FILE and
--frozen-references-sha256 SHA256 to compare.
The digest must be the previously accepted digest of this exact replay
manifest. Neither option is inferred from the local files.
The manifest is JSON with schema_version: 1,
engine: "libreoffice" or "word", and an
artifacts array. Each artifact has path,
bytes and sha256. Paths are canonical relative
paths under tests/corpus/private/refs/<engine>. For
each selected input, include its reference.json,
source.pdf and every ref-*.png. The inventory
must cover exactly the selected documents' required artifacts; subsets,
duplicate paths and extra listed artifacts fail.
Before rendering, replay verifies:
- The trusted manifest digest and every listed artifact's bytes and digest.
- Complete, contiguous page ordinals within the existing page safety limit.
- The source input's bytes and digest against the corpus manifest.
- Schema, engine, DPI, page scope, protection policy and document font substitutions.
- The original receipt's font policy against the current bundled font files and substitutions, and its pinned PDF font inspection for embedded, permitted fonts.
- One original tool/font identity across the batch, with no internal symlink aliases.
The pinned PDF bytes bind the original page count and font
inspection. Replay does not run pdfinfo,
pdffonts, the reference application or its rasterizer; a
new inspection requires the existing reference-generation workflow.
After rendering, all reference artifact hashes are checked again before
reports are published. PNG write/copy failures abort comparison instead
of leaving a successful report with missing raw output.
The JSON report retains schema 3, binary64 scores, DPI 72, the SSIM
threshold 0.95 and every page up to max(ours, reference).
Missing sides still contribute zero. An added
reference_replay field records the manifest digest,
artifact count, original tool/font identity and
reference_tools_executed: false. Exit zero means the
measurement completed; the existing required passing-page rate remains a
separate acceptance check.
Omitting these options retains live reference-tool identity
verification. corpus-tool refs continues to require the
actual reference tools and perform the existing PDF font/page checks
when generating references.
The focused regression suite covers changed inputs/artifacts, incorrect receipt policy, missing or extra inventory members, path traversal, symlink aliases, unembedded/out-of-bundle fonts and changes during rendering. The score round-trip test checks both ordinary and replay reports at adjacent binary64 threshold values.
Validation on 10 October 2026
Candidate b425664b93ac52deea6c59ebe9cb0c261ca12f48
passed 42 focused tests, including nine new replay regressions,
formatting, strict Clippy and normal
corpus-tool/meridian-convert builds on Sagar.
Independent auditing verified all 3280 source files and executable
modes, and all three actual/frozen binary pairs. This is scoped tool
acceptance.
Actual CLI controls also passed in a container without the reference tools: the previous binary exited 2 for missing tools; the new binary exited 2 for a wrong trusted digest and for an omitted digest; an accepted, explicitly separate one-document subset completed with exit 0. Independent auditing checked all four logs, the pinned subset and parent manifests, actual/frozen binaries, the copied reference image, and fully decoded all three control PNGs. The small subset does not establish full-corpus visual fidelity.
The full 383-document measurement was launched separately with the original 575 inputs, 16 exclusions and 3637 pinned artifacts (383 receipts, 383 PDFs and 2871 PNG pages). All 575 input hashes were checked again before execution. Full measurement and independent pixel recomputation remain pending; no new reference PDFs are generated, and the 0.95/60% criterion is unchanged.
The preceding full workspace CI for desktop source
a262f141 remains failed: 2326 tests passed, two failed and
13 were ignored in the printed summaries before the run stopped. Both
failures are Sheets spelling recovery tests which never reached their
controlled worker's held stage. Their original logs and fixtures are
retained; C ABI was not run after that failure. These counts do not
describe a completed successful workspace gate. The previously failing
dictionary writer and actual delayed-Hello first-frame cases passed in
this run, which does not establish the causes of the earlier
failures.
See the machine-readable evidence for source and receipt bindings. Product release and full-plan acceptance remain open.
Nightly G3 on GX10 (17 October 2026)
The owner accepted the LibreOffice 26.8 references of 11 October 2026
as the G3 reference (qa/gates.json). The LibreOffice
26.8.0.3 build that produced them is no longer installed on GX10 or
Sagar (both carry 24.2.7.2), and corpus-tool compare
without replay rejects receipts of another tool version, so the nightly
G3 job (.forgejo/workflows/g3-nightly.yml) replays the
accepted batch.
The replay manifest was not recovered from 10 October (digest
7fc8552c… is not on either server); it was derived by
scripts/run_g3_nightly.py prepare from the batch inventory
recorded by the baseline report of 2082ba21:
common-63fd49f1-ref30-full383-attempt1/batch4/reference-artifacts.json,
SHA-256
da898cee0c1c70b5d4dd242fad4372afcbd02b025307140204c272a81cfbfa94,
3637 artifacts. Every artifact was copied and its bytes and digest
re-checked against that inventory; the copy is read-only. The resulting
manifest has SHA-256
f922423de613ac89bd8ae8e00e3953e9f55309201e712afe34641d858d87bf19
and is pinned in the workflow. The baseline comparison
(cf4c11d0…, the one bound by
docs/integration/evidence/visual-layers-2082ba21-20261011.json)
is copied next to it for the layer deltas.
RU: Ночной G3 сравнивает наш рендер с принятой партией эталонов 11.10
без LibreOffice; манифест выведен из инвентаря партии (sha256
da898cee…), каждый файл перепроверен, закреплённый дайджест
— f922423d….