Datasets:

---

Lexmount

/

LexBench-Web-Fetch

Like 1

Follow

Lexmount 9

This is the separate material follow-up described in #3. It preserves #3's 66 rule corrections in a new cumulative revision; the two PRs still need coordinated review before merging.

Scope: 14 new material/revision/population files, one additive review-index entry, 13 existing validation controls copied unchanged from #3 into definitions/evaluation/scoring-regressions.json, and a synchronized archive/manifest. The 13 controls are checks, not 13 additional benchmark tasks. Existing main-branch review records, task files, scoring algorithm, repository tools, published editions and tags are unchanged.

The cumulative new revision changes only three normalized-text verification records relative to #3: live-041-python-programming-language region 2, live-063-navigation-menu region 5, and live-069-navigation-menu region 5. Source text, ranges and scoring thresholds are unchanged. The same 530 tasks are selected.

Five new source manifests/catalogs bind external executable material packages; raw third-party captures and full reference answers are not uploaded. Obtain a permitted external copy separately; this PR does not provide a public source download service. See definitions/materials/recovered-materials-20261105/README.md for the input boundary and existing import procedure.

Validation of this PR's exact definitions with Browser Eval at 5c36f28756f0659317d2c51e80d0525197df6caa: all five suite definition checks passed after adding the 13 unchanged controls. The archive contains 55 definition/notice files and matches the branch byte for byte. manifest.json binds uploaded source commit 73e882fe05f2700a5ec60496a54efee2c4a000c4; publication commit 55720a1ff7a7e7860f2cfd990b63d0e0c54da7c0 changes only the archive and manifest. Both remote downloads were checked against the validated local files. The default published edition is preserved; the new population file selects the unchanged 530-task set.

Earlier material validation against the proposed combination with #3: all five packages imported, 530 selected cases loaded after normal native preparation, required list references were readable, and changed inputs/overwrites were rejected. All 13 recovered-source controls passed and error controls failed. The three repaired body references passed complete-source controls and rejected error pages. These material checks are distinct from the 13 unchanged registry controls added above.

Remaining review limits: a complete live 530-task evaluation and independent human review have not been completed. Archived controls do not prove that current live pages are unchanged; no Douyu or USPS rule change is proposed here. If the PR branches are integrated or changed during review, preserve both review-index entries and rebuild/recheck the archive against that final source state.

Independent review — changes requested

Reviewed #4 at 55720a1ff7a7e7860f2cfd990b63d0e0c54da7c0 and the linked report.

Please consolidate #3 and #4 into #3. Put the necessary changes in #3, make its archive/manifest match that final state, and close #4 as superseded. Keep one PR for further review.

The requested deliverable is a dataset that another person can download and use for scoring. This review does not request an upload of review dossiers, reviewer transcripts or extra audit packages. Original HTML captures and measured provider outputs must not become prerequisites for scoring this dataset.

1. [P1] Actual reference answers are still missing. The five new manifests name external dataset.json files but do not provide them or pinned download locations. In the proposed 530-task contract, 302 tasks require body-reference text that the public definitions replace with hashes: webfetch 86, dynamic 42, access restrictions 79, articles 95. Directly scoring downloaded static-simple raises ValueError: Empty body region; article-001 raises ValueError: Strict body requirements need every region and a nonempty ending. Supply the final tasks and the necessary reference text/structured answers. A clean public download must load and score all 530 tasks without author-local material. Passing definition checks alone is insufficient. See the material boundary.

2. [P2] Remove unnecessary scoring dependencies and duplicated review history. The new manifests require 980 external files, including 975 beyond the five datasets. Retain only dependencies actually needed for scoring. The new review entry contains 532 records; 529 duplicate #3 and 463 duplicate the existing main-branch predecessor records. Keep concise correction reasons and version/hash references, without copying the full review history again. The current revision tool requires those records, so remove that dependency through the existing owner rather than deleting JSON and breaking export. Preserve already-published editions. See the added review entry.

3. [P1] The scope correction remains inconsistent. Xueqiu article-009 and article-103 drop unrelated download/about/complaint UI, while same-site article-102 still requires 下载App, 关于雪球, 侵权投诉指引 and 登录. Apply one declared scope rule across the relevant tasks, including previous passes; do not restrict corrections to the outcome-selected candidates. Preserve legitimate body text, authorship and other requested content. The changed scope remains a separate task contract, not authority to relabel historical runs as a new matched evaluation.

4. [P1, linked report] Two product lists were parsed incorrectly. Local dynamic-020-webscraper-io-test and dynamic-catalog each return six product blocks with names, prices, descriptions and review counts. The report's parsed records instead use 21 items, 117 items or the preceding 12 reviews as product names, then declare descriptions/reviews missing. For example, the tablet answer contains $69.99, [Lenovo IdeaTab], 7" screen, Android and 12 reviews in one block. The same frozen tasks and outputs pass through the public scorer at the PR's verification commit 5c36f28756f0659317d2c51e80d0525197df6caa. Fix the shared record-boundary parser and rescore the affected family for both methods, preserving old verdicts. Add positive/negative cases for adjacent cards and missing/misassigned fields. New provider calls are unnecessary for this scoring correction.

5. [P2, linked report] Equivalent spacing is rejected. Local article-019 returns ©著作权归作者所有; the required text is © 著作权归作者所有. All other requirements pass, but that space makes the task fail. Auto article-046 returns QCon 全球软件开发大会, reported as missing QCon全球软件开发大会; that task has other failures and must not be flipped solely on this check. Public scorer controls reproduce both false negatives: original spelling passes, equivalent spacing fails, a nearby wrong word fails. Fix shared matching with those positive/negative controls while preserving meaningful English word boundaries. Do not globally strip spaces or add provider exceptions.

6. [P1, linked report] Investigate the example.com reference mismatch. static-simple requires the frozen wording avoid use in operations; both saved answers contain the newer warning about not relying on the domain for testing/monitoring. An independent source fetch on 2026-11-06 returns that newer paragraph too. This is evidence for investigating reference drift, not contemporaneous qualification of the historical calls. Handle source validity across both methods, retain old evidence, and establish independently qualified references before readmission. Do not count missing obsolete wording as proven retrieval loss or rewrite references from provider outputs.

7. [P1, linked report] Bind the scorer and retain the answer fields used for grading. The report cites scorer 931c6786…, but its reproduction page does not identify an obtainable exact implementation. The PR's verification code has scorer e8167a2c…; these are different. article-029 is corrected using a separately returned date, but the downloadable answer and public record metadata omit that field; a review summary stating its value is not a complete verifiable scoring input. Publish the exact scorer revision and retain necessary sanitized answer fields with the report evidence. Keep them separate from dataset reference answers; no extra review dossier is requested.

What the low Auto score does and does not show

The saved judgments aggregate correctly to local 390/530 and Auto 113/530. However, 38 Auto article failures pass an applicable body-coverage check, and 26 fail only the records/required-text check. Those are not all missing-body failures; this does not establish that all 26 should pass. Some required metadata may still be legitimate.

The API request format is correct: POST /v1/extract, with extract.options.engine_preference: auto. Auto also has real delivery failures: the quote-list answer contains Loading/footer text and zero of ten required records, and another response is a captcha page. Its low score mixes actual delivery gaps, task scope and scoring defects; it cannot be read as body-retrieval success alone. There is no evidence here establishing intentional cheating.

Checked scope: the complete 1,060-row matrix and all 557 saved machine failures. Public-code replay completed without errors; two failures pass on that code, with no body-decision changes. That code is not the report's missing exact scorer, so this is not an authoritative full rescore. The 503 machine passes were not independently semantically reviewed. No new provider calls were made. The 55 archive files and 65 manifest entries match the branch; all five definition checks pass; the three text-binding repairs recompute correctly. These checks do not supply the missing reference answers.

Please address these concrete issues in the consolidated #3. Acceptance should be a clean download that can score the declared tasks, plus corrected report judgments where the scorer is wrong—not more uploaded review paperwork.