Datasets:

---

Lexmount

/

LexBench-Web-Fetch

Like 1

Follow

Lexmount 9

Adds an independently source-bound annotation cohort with main/full acceptance profiles. Existing task suites and historical annotations are unchanged. Full reference bodies remain in the matching approved input dataset; public definitions retain hashes. No provider responses, measured results or credentials are included. Scoring and report support: https://github.com/lexmount/browser-eval/pull/268

Review of refs/pr/6: this isn't ready to merge. Its parent is the current main, f01d834, but the commit rewrites README.md and manifest.json from an older generator. That reverts #3 (46d9107) and f01d834.

1. P1: the edition block is deleted from manifest.json. It held v2.1.0, 550 tasks, population_sha256 and the configurations. As a result, make check fails:

   • tools/rebuild_publication.py:30 raises ValueError: The existing edition metadata must be retained.
   • tools/test_review_scope.py:39 raises FileNotFoundError …/definitions/materials/primary-content-20261103/content-scopes-530/dataset.json, because the new registry suite has no materials directory.

   On main, make check passes ({"task_count": 550, "archive_members": 61}).

2. P2: the README loses existing content.

   • The diff is −118/+24. It drops the "Short name: LB-WF" line, the License paragraph saying the definitions/materials/ reference texts are third-party and not Apache-2.0 (they are still in the archive), the "Dataset versions" section and all of the 530-task revision and selection docs from #3.
   • "1,080 task entries" double-counts. The content-scopes-530 IDs are exactly the 530 tasks already in populations/primary-content-20261103.json, so this file adds scoring profiles, not tasks.

3. P2: it points at code that isn't on Browser Eval main. manifest.json tooling.revision and the README citation link pin 691cd74a, an unmerged commit of browser-eval#268 whose CI is failing. annotation-provenance.json also names 7 regression_ids that exist only on that branch. If #268 is amended or squashed, the dataset will point at a commit and controls that don't exist on main.

4. P3: two tasks disagree with their own full profile. In definitions/evaluation/content-scopes-530/tasks.json, for webfetch--static-simple and webfetch--redirect, the top-level body_requirements / required_facts differ from the full profile. A run without a scope would execute against the top-level rules but score with full.

Checks that passed:

• The new files' hashes are self-consistent across the manifest, the registry and the archive.
• The file has 530 unique IDs, matching the 530-task population.
• With the #268 code, load_dataset loads all 530 under main, under full and with no scope.
• Nothing leaks: no reference bodies, local paths, provider outputs or latency data.

Suggested order:

1. Fix and merge browser-eval#268.
2. Rebuild this PR on current main with the repository tooling. Keep the edition, the license text and the existing README sections, and pin the merged commit.
3. Release it as a minor dataset version, because it adds scoring profiles to the same tasks.

v2.2.0 is being released for the #3 reference corrections, so this would become v2.3.0.

P0/P1 re-review of refs/pr/6, head 9ddbdad2, unchanged since the last review. Dataset main is now c12559d2 (tag v2.2.0). This PR has three P1 issues.

1. P1: it conflicts with v2.2.0 and would revert that release.
   • git merge-tree --write-tree main pr6 reports content conflicts in README.md, manifest.json and definitions.tar.gz.
   • The PR's archive (5117401…) still carries the old VERSIONING.md without the v2.2.0 row, and its README and manifest predate v2.2.0.
   • Resolving the conflicts in this PR's favour, or re-uploading its files, would remove the v2.2.0 edition from main.
   • Apart from VERSIONING.md, the definitions are otherwise identical, so a rebuild on main only needs this PR's content-scopes-530/ and registry.json changes.
2. P1: deleting the manifest edition block makes make check fail.
   • tools/rebuild_publication.py --check raises ValueError: The existing edition metadata must be retained at tools/rebuild_publication.py:30.
3. P1: the new registry suite breaks the dataset's own test.
   • definitions/registry.json:61-69 registers content-scopes-530.
   • tools/test_review_scope.py:19-37 then opens definitions/materials/primary-content-20261103/content-scopes-530/dataset.json, which isn't in the archive: ERROR: setUpClass … FileNotFoundError.
   • For comparison, main c12559d2 passes make check (31 tests, {"task_count": 550, "archive_members": 61}).

Still open below P1:

• tooling.revision and the README links (lines 128, 181, 188, 198, 219, 225, 240 and 244) pin 691cd74a, the first commit of browser-eval#268. That PR's head is now a59d90d4, so the pin is stale.
• README line 82 still says "1,080 task entries", which counts the same 530 tasks twice.

What works:

• With browser-eval#268 at a59d90d4, all 530 tasks load under main, under full and with no scope, once "tool": "web_fetch" is added (the other raw tasks.json files omit it too).
• I found no new private-reference leak.

Fix: rebuild on current main with tools/rebuild_publication.py, keeping the edition, README and VERSIONING content. Give the new suite its materials or exclude it from suite_dirs(). Pin the merged #268 commit. Publish it as v2.3.0.

Updated at 048b9962c11ee8518374133203275c306bedb888. Dataset main c12559d2 (v2.2.0) is now an ancestor of this PR, and the archive/manifest conflicts are resolved.

The 550-task edition object, existing README sections, license notices, VERSIONING.md, task definitions, reference materials and viewer configurations are preserved. The redundant content-scopes viewer has been removed: this cohort annotates 530 existing tasks rather than adding another 530 tasks. Historical revision tests now use the population’s declared parent suites, so they do not request nonexistent materials for the annotation cohort.

Validation: 33 tests pass; publication check reports task_count=550, archive_members=65. The manifest binds uploaded source 5f15e1e4cf6477b81395fee89ffeecb01fd5cdfb.

The cohort registry records its unreleased implementation dependency on browser-eval#268 at a59d90d491091983a919464cc9e84924d8a9fb9f, instead of replacing the released edition’s tooling pin. That cohort pin must be updated to the merged code commit before an annotation release. No release version was changed and the PR remains open.

Re-review of head 048b9962 against main 1948cae5 (the Webstone rename).

Fixed since the last review:

• It merges cleanly and no longer reverts v2.2.0.
• The manifest edition is kept: v2.2.0, 550 tasks.
• suite_dirs() is limited to the population, and make check passes on the PR head (33 tests).
• "1,080 task entries" is gone.
• The static-simple/redirect base fields are now documented as the historical task.

Still blocking:

1. P1: the merge result fails make check.
   • Main 1948cae5 edited THIRD_PARTY.md and VERSIONING.md but didn't rebuild the archive or manifest that bind them. So main itself fails tools/rebuild_publication.py --check with Changed publication file: THIRD_PARTY.md, and this PR inherits the failure.
   • This PR's archive (f864f077) also still carries the pre-rename notices.
   • Fix: rebuild on 1948cae5 with tools/rebuild_publication.py --source-revision <uploaded commit>. That also replaces the P3 source_revision 5f15e1e4, a superseded PR commit whose tree still had data/content-scopes-530.jsonl.
2. P2: the pin still points at the wrong commit.
   • definitions/registry.json:70-75 and README.md:359-361 pin a59d90d4, but #268 was squash-merged as 1287fdf8. a59d90d4 is not on Webstone main.
   • Fix: pin 1287fdf8 and drop the "must be pinned to the merged commit" wording.
3. P2: it uses the old repository names.
   • registry.json:72 and README.md:359 still say lexmount/browser-eval / "Browser Eval".
   • Fix: use lexmount/webstone / Webstone.

Minor:

• P3: VERSIONING.md doesn't mention the added annotation cohort yet. That belongs in the next release row.
• P3: registry.json reorders the keys of the five existing suites, which adds about 70 lines of noise.

Checks that passed:

• All 530 tasks load with Webstone main 1287fdf8 under main, under full and with no scope.
• validate() passes for all 6 registry suites.
• All 7 regression_ids exist in Webstone main's scoring-regressions.json.
• No private reference body leaks.

I'll push these fixes to this PR and merge once make check passes on the merge result.