[lanyue-llk]

Problem and result

Live and saved responses could receive different verdicts, and a task revision could name regression controls without executing them. This PR gives Browser Eval and other callers the same deterministic scoring and revision-export contract.

Changes

• Use the public scoring orchestrator and task-result predicate for live and saved responses. Keep binary acceptance separate from text-quality diagnostics.
• Reject incomplete body facts, word fragments, unrelated destinations, fields combined across separate records and invalid file evidence. Accept equivalent link and Markdown serialization when the required content is present. Resource aliases allow only declared asset-preserving parameters.
• Preserve frozen task definitions and historical dataset identities. Explicit revision exports verify the parent, changed leaf fields, independent source hashes and executable positive/negative controls before producing a new identity. inspect-revision exposes those checks.
• Verify retained reference answers against the same scorer and reject accepted error pages. Missing original page bodies remain visible in inspection; a hash alone does not establish complete source coverage.
• Keep revision JSON and scoring-regression JSON as installed, executable benchmark inputs. Report instances, provider outputs and publication targets stay outside this package.
• Document contribution and title conventions, and update the shared banner in both README editions.
• Use Apache 2.0 for project-owned code and definitions, as explicitly requested by the repository owner. Both source and wheel distributions ship the canonical license, LexMount copyright notice and third-party notices.
• Check English maintainer identifiers, comments and docstrings in the installed package, while preserving original-language task evidence.

The branch includes current main's article acceptance contracts. Expanded main-content references and unified mixed requirements remain in draft #8. Execution, retained evidence and all three report pipelines are handled by Browser Eval #56.

Validation

• make check: 2,541 tests and 321 subtests passed after integrating current main.
• Python 3.11–3.14 CI passes at 6bd263ed; the PR has no merge conflict.
• Revision inspection executes registered controls and rejects missing review records, incorrect field differences, unbound sources, empty reasons, one-direction controls and failing checks. Seven of ten historical corrections have executable retained-source controls; the other three retain source hashes without the original page body.
• The harness pins this exact commit. Its retained real-data report instances cover Web Fetch (3,024 results), Browser Use (180) and Browser Trajectory Replay (450).

Historical report scope

The published 430-task report retains its frozen snapshot and scorer binding. Its complete comparison against the audited e1b6efc1 definitions contains 286 matching contracts, 91 retrieval-hint-only differences and 53 other differences. Original/revised-definition pass counts are 378/360/229/223/217/190 and 385/367/233/227/222/194 respectively. LZNOIP's eight canonical English navigation aliases are preserved by the report's bound revision. These saved results are not recomputed when this package's source identity changes.

The full scope and source-copy audit are attached to the Notion review. The existing report keeps its normal version presentation and permissions.

Scoring and partial reruns

Reproduction and canonical-contract comparison

The historical scorer is preserved at report-scorer-20261022, paired with harness report-renderer-20261022. The full frozen task snapshot, its parent, both annotation revisions, canonical contracts and verdict inventories are attached to the v8 review. Report data remains outside Git and retains its existing private access.

Both offline comparisons use the same 430 case IDs and 2,580 hash-verified saved responses, including the previously declared eight recovered originals and 199 public projections. Scorer 6bd263ed reproduces all published verdicts with zero changes. Applying the canonical base contracts at e1b6efc1 gives:

Configuration	Report snapshot + revisions	Canonical base contracts	Pass → fail	Fail → pass
Moli cumulative	385/430	366/430	20	1
Moli 1.1.15	367/430	349/430	19	1
Chromium	233/430	219/430	14	0
Lightpanda	227/430	215/430	14	2
Obscura	222/430	209/430	14	1
Kitesurf	194/430	183/430	11	0

All 97 verdict differences occur in 27 of the 53 tasks with contract/reference differences. The 286 identical tasks and 91 retrieval-hint-only tasks have no verdict differences. These canonical scores are a diagnostic under different task requirements; no new provider calls were made.

The differences include concrete historical reference corrections: the canonical retail task requires “Returning & Exchanging Online Orders” while the report requires the observed “Returning or Exchanging Online Orders”; the canonical Arch task combines navigation and heading into “Arch Linux Official Package Search” while the corrected heading is “Package Search”. The canonical example-page task also requires the title and a different body region; the saved Moli output reaches 66.67% against that region, below its 80% requirement. Changing to those base contracts would change the benchmark definition, so this comparison does not replace the declared report snapshot.