Datasets:

---

Lexmount

/

LexBench-Web-Fetch

Like 1

Follow

Lexmount 9

This PR proposes one rule revision for a fixed 530-task selection. It revises scope in 69 cases (37 articles, 14 access-restriction tasks, six dynamic tasks and 12 popular-site tasks). The rules make unrelated site controls optional and retain required article content, tables, links, records and displayed comments.

The revision covers three additional Xueqiu articles (article-044, article-102 and article-104) and incorporates the three reference verification repairs from PR #4 (live-041, live-063 and live-069). Those three repairs preserve the original reference ranges and scoring thresholds. Rationale and evidence

The selection keeps 130 general web tasks and 100 tasks from each of the other four suites. It follows the original report population. The ten excluded dynamic examples duplicate selected entries. Three of the ten excluded challenge examples share URLs with selected entries. The selection was made after observing outcomes.

All 27 dataset checks passed locally. A separate local package exercised scoring for all 530 tasks with matching reference inputs. The public repository provides definitions and rules. Full scoring requires separately obtained reference inputs, so the review request for a self-contained public scoring download remains open. The rules and population are on the source commit. Synchronization of the public archive and manifest is pending.

Treat this proposal as a separate task contract for future matched evaluations. The review used automated checks and saved-answer controls. Independent human sampling and formal adoption remain pending.

Reviewed at 3a1f5a96, alongside https://github.com/lexmount/browser-eval/pull/131 (first checked with its code at ba87b8b; this PR is unchanged since). I recommend not merging yet. Findings 1 and 2 need a decision. The others are smaller fixes.

Confirmed: the new and changed definition files, VERSIONING.md and the archive match their manifest.json hashes, as do all 44 archive files the manifest names. Applying the revision with #131's apply_annotation_revision changes exactly 64 tasks relative to review-550-20261023, only task_prompt and acceptance_contract, and the set matches the rationale file. Viewer data and the default catalog are unchanged.

1. The same requirement now depends on whether every method failed. Only tasks with 11 failures were reviewed, so the rule stops at that boundary:

• 244 of the 459 unreviewed tasks still require at least one string this revision removes as page shell (127 require five or more).
• article-103 and article-009 are both Xueqiu articles with the same predecessor prompt. 下载App, 侵权投诉指引 and 关于雪球 are removed from 103 and still required by 009.
• article-091 and article-064 are both Quanta articles with the same predecessor prompt, and 21 strings are removed from 091 only.

The 244 is an upper bound, because some of these strings may be legitimate targets elsewhere, such as a homepage task. The same-site pairs are the clearest cases. Outcomes selecting candidates is disclosed, but here they also decide which tasks get the corrected contract. The Browser Eval rule is that an acceptance correction "must fix the shared cause" (AGENTS.md:65). Suggested fix: write the shared rule, for example that article tasks do not require site shell, and apply it to every task in scope whatever its outcome, or keep this revision as a clearly labelled diagnostic and out of any ranking. Please also publish before/after pass counts per method, since a task-level rationale doesn't show who gains.

2. By the project's own rules, this is a major version change. VERSIONING.md classes changed requests or required output scope as Major. create_release agrees: on the articles suite it raises at 2.2.0 ("Changed task population or input contract requires a major version"), and at 3.0.0 compare_releases returns separate. That conflicts with rescoring saved outputs for a method comparison. If the prompts stay unchanged, this could be a minor acceptance correction, and I found no provider adapter that reads task_prompt. That looks like the simplest way out.

3. The prompts now contain revision history. All 64 gain 本次范围修订仅取消以下页面外壳文字的必需性：…, adding up to 2,321 characters. They also say 以该案例原有要求为准, which points readers to a superseded contract. A task prompt should describe the task. This explanation belongs in the rationale file.

4. The manifest drops edition while the docs still rely on it. VERSIONING.md, which this same manifest binds, says manifest.json records "the edition's active population" and defines its population_sha256, but the block is gone. Browser Eval's publication.py doesn't appear to emit it, so any rebuild will drop it again. This is a metadata change the description doesn't mention. Please teach the builder to carry it, or update the docs.

5. source_revision doesn't resolve. The manifest now names this repository with 5f0764dc143046be055b7a4e4ec9276372d7c9ca. The HF revision API returns 404 for it (200 for 3a1f5a96 and cab10346), and it isn't on lexmount/lexbench-webfetch either. It may be a local checkout commit, since upload_folder creates a new one. pinning.py copies this value into the data lock, so the dangling reference would spread.

6. The "independent page evidence" can't be checked publicly. No review-evidence/ files are in the manifest or archive, only their hashes. Three generic reasons cover 1,571 of 2,471 rule decisions. Also, the card says export "fails on missing or altered bindings", but --review-source is optional, and without it export skips those files and records them as unavailable. If the captures can't be redistributed, please publish capture date, method and reviewer for each task, and correct that sentence.

7. Merge order and controls. On current Browser Eval main, this revision raises Unsupported annotation revision, so the card's instructions only work once #131 is merged. Please name the order: #131, this PR, then a separate pin. The two new controls are both article-specific (article-primary-scope-*), while 32 of the 64 changes are in other suites.

Limits of this review: I didn't have the captured pages, so I can't say whether individual removals are right. Some look well founded, for example dynamic-096, whose frozen prompt already excludes navigation while its 40 anchors required it. I also didn't review all 64 cases by hand.

Update after a second review: I rescored saved outputs from an older private 430-task run (six browser configurations, different dates and methods from the report). 15 of the 32 revised tasks with outputs there already passed for at least one configuration before any edit, and only 4 of 192 verdicts flip with the edits. So "all 11 methods failed" doesn't by itself show that a requirement was wrong, and page shell may not be the main cause. The articles suite isn't covered.

Status on #131: the author's reply at a488cf6a (https://github.com/lexmount/browser-eval/pull/131) confirms the 244-task and article-103/article-009 findings, the missing edition block and the 404 source_revision, and withdraws this revision as authority for a primary ranking. On the report's own results, 78 of 640 affected verdicts change from fail to pass and 547 still fail. I haven't independently checked those figures. If this PR stays open, findings 1 to 7 still apply to any future release of it.

Second review, at 4b3a54ce, using the code from https://github.com/lexmount/browser-eval/pull/131 at 1c84ad26. I recommend not merging yet, mainly because of items 1 and 2.

Confirmed fixed: the edition block is back. source_revision (e11298fa) resolves. The archive matches its stated SHA-256, and all 47 manifest files inside it match. Revision history is gone from the prompts. Applying the revision changes exactly 64 tasks, only task_prompt and acceptance_contract. The population has 530 tasks. All 110 regression IDs resolve, with no ID collisions against the PR or main, and no existing control lines were removed. Calling this a separate major edition for fresh runs is also the right framing.

I also checked whether the removals left tasks too easy to pass. The tasks with the fewest text anchors left (dynamic-096 40→1, dynamic-072 5→1, article-003 6→2) still have link-record or body-coverage checks, so I found no problem there.

1. P1. A fresh run doesn't fix the rule inconsistency. The review notes accept this finding and leave it out of scope. But the stated use is now formal evaluation, which makes it matter more:

• 252 of the 466 unrevised tasks in the 530 population still require at least one string this revision removes. 127 require five or more. Among unrevised articles, that is 55 of 68. These are upper bounds, because a string can be legitimate on another page.
• The same-site pairs remain. article-103 no longer requires 28 strings such as 下载App and 关于雪球, and article-009, also on Xueqiu, still requires all 28. article-091 and article-064 on Quanta differ by 21.

So whether site shell is required depends on whether all 11 methods failed that task in an earlier run. Fresh runs inherit that. It also affects methods unevenly: a method that strips navigation can pass a revised task and still fail its same-site twin. The #131 reply says Firecrawl's selected reruns turned off onlyMainContent, which fits a benchmark that rewards returning page shell, though it doesn't prove it. Suggested fix: write the rule as a scope statement and apply it to all tasks in a suite regardless of outcome, including tasks that already pass. If that's too large for now, keep this as a labelled sensitivity edition and don't use it for formal rankings.

2. P1. The 530-task contract hasn't been shown to run end to end. Your description and review notes say full export remains unverified for webfetch, popular sites, dynamic content and access restrictions, and that an auxiliary source and other frozen inputs are still missing. That covers 430 of the 530 tasks. Please make this a merge condition, because the point of this PR is to define something others can run.

Related: nothing in Browser Eval at #131's head reads definitions/populations/, and scripts/run_certified_benchmark.py has no revision or population option. The 530 selection only reaches a run through a dataset-local tool and a hand-passed --case-ids. Please show one small run through the certified runner, with the population hash bound in its output, so a run on a different task set can't carry this edition's name.

3. P2. The controls don't test what the revision changes. A complete source answer passing and an access-denied page failing are useful checks, but a complete page contains the shell strings, so I'd expect it to pass the old rules too. An error page would fail either way. I found no control for the actual claim: a clean main-content answer that fails the old contract and passes the new one, paired with a shell-only answer that fails both. Also, the provenance file has full_answer_control on 41 of the 64 reviews, while the description says all 64 are covered. Please say where the other 23 are recorded.

4. P2. "Reviewed" needs a plainer meaning. In the provenance file, capture_reviewer is null for all 64, publicly_replayable is false for all 64, browser_visibility_verified is false for all 64 current observations, and review_actor is assistant for 59 reviews. This is all disclosed in the file, which is good, but the PR title and description only say "reviewed". Please say in the description that the review was automated and can't be replayed publicly, and ask someone other than the author to sample some removals before this becomes a formal edition.

5. P2. The 20 excluded tasks have no stated criterion. Their reason reads "Outside the requested 530-task report population", which says who asked but not why. The #131 thread calls this a post-outcome decision. Please record the rule, so a future reader can tell why those 20 are out.

Merge order: #131 is currently conflicted with main, including in scorer provenance, and its scorer will change when that's resolved. Please re-run the final checks against whichever scorer gets pinned.

Not checked: I don't have the review-evidence files or reference inputs, so I can't confirm individual removals or the case-level control results. I didn't re-run my earlier saved-answer rescoring on this final contract.

Please consolidate #3 and #4 into this PR (#3), rather than maintaining two overlapping cumulative revisions. Bring #4's necessary reference-answer and three binding-repair changes into #3, update its archive/manifest to the final combined state, and close #4 as superseded. Further review should use one PR and one final definition state.

The independent review on #4 identifies the remaining missing reference answers, duplicated review records, inconsistent task scope, and concrete scoring problems in the linked report. Please address those in the consolidated work. The dataset needs the actual reference answers and fields required for scoring, not uploaded review dossiers, raw HTML archives, or measured provider outputs. Report scorer/output corrections belong in the existing scorer/report owner.

The acceptance condition is straightforward: someone should be able to download the public dataset and score all declared 530 tasks without private material or author-local caches. Keep the brief correction rationale and necessary version/hash bindings; avoid copying the full review history again.

Follow-up: all 503 saved passes reviewed

I have now checked the remaining 503 reported passes (390 local, 113 Auto), covering the full 1,060-result matrix together with the earlier 557-failure review. The report reviewed is still version ver_4a46501912b44a80bc00dbd0acdf4fe2.

This was an independent code-and-output review, not a blind human sign-off. It included replaying every pass with the public verification scorer, checking actual delivered file bytes, and investigating suspicious labels, short/empty outputs, titles, comments, structured records and replay disagreements. No new provider attempts were made. I am reporting findings here, not requesting an upload of review dossiers or extra audit packages.

New finding 1 — [P1] Named-link scoring accepts buttons, ratings and icons as entity names

The report's parsed records contain concrete examples:

• Local dynamic-087-joyn-filme: the parsed name for /filme/godzilla-bpcsu8vy4pwf is 12, an age rating, although the answer also contains the correctly linked heading Godzilla.
• Local dynamic-091-pt-br-kabum-hardware: some parsed product names are the private-use favorite icon U+E87D or rating text, rather than the returned product name.
• Local global-homepage-035-fandom-com: the TV entry is named View All, although a separate TV link to the same destination is present.
• Local china-homepage-075-volcengine-com: product names are recorded as 查看详情, despite adjacent links containing names such as 豆包大模型.

This is more than a display issue. With the same frozen task contracts, the public scorer accepts these negative controls:

These are synthetic rejection controls, not substituted benchmark outputs. The exact report scorer is still not bound to obtainable code; the controls execute the public verification scorer, while the report's own saved parsed records demonstrate the label-selection defect.

Do not flip those four historical passes merely because their parsed names are wrong. Their original answers contain meaningful labels for all 76, 67, 3 and 19 counted destinations respectively, exceeding their required quotas. I checked the alternative labels per destination. The new finding is a false-positive acceptance path and inaccurate parsed evidence; it is not proof that these particular original answers lack names.

Required change: fix the shared named-link selection/validation. Choose the actual entity title from its declared field/card/heading and retain a meaningful label before deduplicating destinations. Generic actions, age badges and icons alone must not satisfy the name requirement. Include both multiple-label positives and generic-label-only negatives. Do not blindly reject every numeric title: a genuine title such as 1917 must remain valid when it is actually the title. Audit this shared behavior across the named-link family, not only these examples. Preserve original outputs and rescore affected judgments consistently.

New finding 2 — [P2] A reported comment count is not supported by the visible comment boundaries

Local article-032 reports 114 comments. Its saved Markdown contains 24 Collapse Expand comment blocks under Top comments, including genuine author/message content. The output also repeats author profile material. The displayed count needs to be recomputed from genuine comment records, excluding profile biographies and other surrounding blocks. A page's global “189 comments” label is not evidence that all those comments were returned.

The task requires 20 comments, so this count discrepancy alone is not a reason to mark the original answer failed: the saved answer contains enough genuine comment blocks. Fix the diagnostic count and add controls that prevent profile blocks from inflating the comment quota.

Pass-review decisions and limits

• 471/503 passes reproduce directly with the public verification code. 32 differ: 27 Auto and five local. These are version/input-surface disagreements, not 32 demonstrated false passes. They concern 21 title checks, 12 records/required-text checks, two comment checks and two body checks; categories overlap.
• The saved titles, named-link fields, actual comment text and equivalent typography explain the substantive disagreements I could check. For example, the Google Sign in and Weibo category labels are explicitly present in returned links; the 20 GraphQL reviews are present with dates and text. It would be incorrect to require them to be duplicated in main_text merely to satisfy the older parser.
• Auto article-029 remains an evidence gap: its required October 2024 value is asserted in a review summary as result.extra.published_at, but the downloadable answer and public record metadata omit that actual answer field. I cannot independently confirm that one corrected pass from the public scoring input. This is an evidence defect, not proof that the provider omitted the date. Retain the necessary sanitized answer field with the report evidence; do not put provider outputs into the dataset.
• The three apparently empty local outputs are legitimate file-delivery results: the PNG, WebM and MP4 binaries are present and match the expected signatures, sizes and SHA-256 hashes. They are not empty-body false passes.
• The short Scrapy answer is also not a demonstrated omission: its actual README prose is short; the larger local answer mainly includes GitHub page surroundings. Likewise, article-008 was explicitly qualified as the displayed portion, so its subscription notice alone does not prove failure under that scoped task.
• I found no additional confirmed whole-task false pass in the original 503 answers. This does not cure the named-link rejection-control failure, the missing date evidence, or the exact-scorer reproducibility gap already reported.

Which report conclusions hold, and which must change

Supported: the saved matrix has 530 tasks per method; its saved judgments sum to 390/530 local and 113/530 Auto. The API request uses the documented endpoint/options. Auto has genuine missing-list, captcha and missing-body examples. The local route delivered the three tested file bytes. Under the report's existing saved judgments, local completes more declared contracts.

Must change: the two shifted product-list failures and spacing false negatives from the earlier review; the inaccurate named-link/comment diagnostics and permissive name gate above; missing reference answers and inconsistent task scope in the consolidated dataset; and missing exact scorer/answer fields needed to reproduce the final scores. The 38 Auto article failures that pass applicable body coverage, including 26 failing only required-text/record checks, prevent interpreting 21.3% as body-retrieval success. A pass does not establish clean output, complete article text beyond a qualified displayed scope, semantic truth/freshness of changing records, or media understanding. Historical final-call latency values are not a matched same-window speed experiment.

Please keep #3 as the consolidated PR and address the dataset issues there. Scorer/report fixes belong in the existing shared owner. Do not publish an invented corrected overall percentage or change stored verdicts one by one; fix the shared causes and recompute from retained evidence.