Derive rationale changes from the data and apply one body-occurrence shell rule
lanyue-llk
6b3e1e5
Derive each case's removed requirements from the parent and revised tasks. Add
explicit, hash-bound entries for the live-041, live-063 and live-069 binding
repairs. Apply one rule to the Xueqiu and Quanta articles: a shell label stays
required only when the frozen article body contains it, so 044 keeps 更多, 102
keeps 全球, 064 drops Computer Science and 091 keeps Physics, Mathematics and Topics.
Remove proposal wording, fix the score --metadata example, state the license
scope and the private inputs export needs, and re-materialize every suite.
Inherit unchanged review entries from the predecessor revision
lanyue-llk
f92a931
Publish the articles-v11 reference texts for the 530-task revision
lanyue-llk
8bbbce3
Add the articles suite under definitions/materials/primary-content-20261103 from the
hash-matched articles-v11 dataset (PR #5), bind its resolved cases in the population,
and record the article control results. All 530 selected tasks are now scorable from
the published archive.
Publish reference texts for 430 tasks and unify the Xueqiu shell rule
lanyue-llk
ad9054a
Add the resolved, scorable datasets for webfetch, dynamic-content, popular-sites and
access-restrictions under definitions/materials/primary-content-20261103 with
scoring-only schema-2 manifests, and bind their cases in the 530-task population.
Apply one declared shell rule to all six Xueqiu article tasks. Merge the oversized
rationale and provenance copies into one compact per-case rationale. Update the
README and THIRD_PARTY notices for the included reference texts.
5 days ago
Bind the 530-task selection to the consolidated rules
Extend the original 64 changes only to article-009 and article-064. Preserve substantive body/comment/source requirements and all merged editions. Rechecked all 66 cases against the current scorer: full answers and cleaned answers pass, shell-only and error answers fail. Record exact revision hash for the companion runner; 530 selected identities unchanged.
Add before-after controls for all 64 cases and fix Vuetify navigation
Record source-derived scoring controls, complete all full-answer indexes, disclose automated review limits and explain the unchanged 530-task selection. Remove only residual navigation labels in the two already-scoped Vuetify cases; preserve frozen editions and substantive requirements.
Complete source controls for the remaining four reviewed cases
Add five archived source bindings and record complete positive/negative checks for 60 of the original 64 changed cases. Four popular-site cases and full-suite input export remain unverified. Preserve all task rules and historical editions.
Verify revised article source answers and fix citation spacing
Adds the opt-in primary-content-20261103 task scope revision supporting https://github.com/lexmount/browser-eval/pull/131.
The review examined 91 suite-qualified tasks with saved failures for all 11 methods. Exactly 64 tasks change relative to review-550-20261023; 27 retain their requirements. The rationale file lists each change and retention reason. Results identified candidates for inspection; task purpose and independent page evidence determined the revised requirements. This is a disclosed post-outcome scope review.
Original frozen inputs, the default 550-task catalog, existing revisions and release tags remain unchanged. The 530-task report is a subset, not a dataset deletion. The cumulative revision preserves prior corrections. No provider responses, measured scores, credentials or captured pages are included.
Validation: all 550 tasks pass the maintained input checks; exactly 64 prompt/acceptance changes against the predecessor; parent hashes and review controls verified. Th…
14 days ago
Preserve current definitions from 5a5b9367118dcc25f126ee16ef4de2d7c666f190
