definitions/evaluation/articles/README.md

Article suite authoring

The registered articles suite (articles-v11) contains 100 qualified URLs: 95 fixed articles and five changing post collections. cohort.json preserves candidate identities and revision history; rules.json contains all 100 verified contracts. Public tasks, the catalog and input manifest bind the complete executable dataset. Full reference bodies and source observations remain external.

Candidate identities and source revisions are recorded in cohort.json; gaps in case numbering preserve the original candidate IDs. Earlier authoring inputs remain immutable. Pre-release version-1 runtime contracts are unsupported. Earlier frozen input bytes are preserved; published results must identify their scorer version.

Four Weixin cases use user-supplied webarchives and an explicit initial-load, no-interaction attestation. The archives, source markup and full body references remain external and hash-bound. This provenance does not claim independent live browser access. They replace four singleton-platform cases without comment contracts. Retained folded pages have explicit complete-displayed-excerpt scope; they do not claim the hidden continuation of the original article.

Existing contracts used as references

The popular-site tasks separate stable site labels from changing named links (for example, global-homepage-013-yahoo-co-jp). The dynamic-content tasks add independent source counts, same-record fields and full-body regions (for example, dynamic-062-zh-sspai-home and dynamic-023-en-mui-data-grid-pagination). Follow the source-review guidance for ordinary visibility, complete scope and independent evidence.

Reuse those scoring contracts where applicable, but do not copy their current items or counts. A named-link floor alone cannot establish article completeness: a fixed-article case must also require the qualified full article text. Changing collections use the separately disclosed post-coverage contract. Likewise, an excerpt directory cannot become an article merely because a list contract can count its links. Existing frozen suites and input versions remain unchanged.

Required scope

Open the fixed URL and observe its first automatic load, including ordinary already-loaded content below the viewport. Do not expand, switch tabs within the page, log in, paginate, accept a wall or scroll to trigger loading. Reference authors may close a login, subscription or similar obstruction through its ordinary dismissal control. Preserve before/after evidence and verify that the article was already loaded and its text did not change; do not use dismissal to unlock or load additional content. This is a reference-review exception only. Evaluated Web Fetch tasks still permit no browser clicks, login, expansion or other content-loading interactions. The historical reference must contain the complete qualified displayed article/post body, not a selection of convenient paragraphs. Keep headings, paragraphs, textual lists, table text, captions and ordinary body links within their declared scope. Exclude image lettering, alt text, OCR, hidden markup and collapsed content. For retained cases with folded prose, disclose a displayed-excerpt scope and require all ordinarily displayed text within that scope. Preserve the visible cutoff and expansion control as evidence; do not claim the full original article or substitute hidden DOM text for the displayed excerpt. A clipping or fading boundary needs rendered evidence, not an assumption based on accessible text.

For article feeds, require actual post text in addition to record identity, same-record fields, real owned links where present, and independently observed record counts. A bare directory or search-snippet list is not an article. Use a separately reviewed post_collection contract for changing feeds; a fixed-body contract cannot freeze their membership. Do not silently drop difficult platforms. Record authorized replacements or removals in cohort history, with the original entries and review basis.

Stable and changing regions

Review every relevant ordinary text region: title, attribution, navigation, sidebar, footer, controls and the article itself. There is no phrase quota. Each chosen group needs an original-language phrase, location and reason to expect stability. Do not freeze current news recommendations, comments, reaction counts, ranks, personalized account state or current copyright years. Historical numbers inside a fixed article remain part of its complete body. Consent prompts depend on session state and are not mandatory site text. Record any obstruction and any permitted dismissal when reviewing the article scope; do not accept binding terms or purchase access to obtain the reference. A nonmodal notice alone does not establish a reading gate: inspect the article's actual visibility, clipping, inert state and any modal backdrop. Already-loaded ordinary document text below the viewport remains in scope; hidden or collapsed body text does not. Each declared stable group must have a matching review ID, source region, stability basis and observation date. Repeated basis/date fields may be stored in per-case stable_text_review_defaults; explicit row values take precedence. IDs and source regions remain explicit, and empty overrides are invalid. Page-identity patterns are regular expressions; escape punctuation when the intended identity is literal text.

Declared stable groups are mandatory in the current fixed-article contracts. They supplement distributed body coverage; title or navigation matches cannot replace missing article text. Reviewers divide every source line into contiguous natural chapters, paragraphs or quotations with readable labels and explicit one-based inclusive ranges. Current segmentation version 3 allows at most five regions. Merge adjacent reviewed components at their existing boundaries, preserving complete text and order. Legacy segmentation versions remain readable for prior frozen inputs. Automatic paragraph detection is insufficient for arbitrary DOM line breaks. Each region must contain enough ordinary text to yield a six-character fragment; merge shorter labels into their adjacent natural component. The version-4 body contract requires every region to reach 80% fragment coverage and independently requires 80% coverage of the reviewed reference_scope.end. A missing or empty ending is invalid. The compiler adds this requirement only to fixed articles; changing post collections retain their separate contracts. The generic scorer retains legacy version-3 behavior for older frozen inputs. These are fragment-coverage tolerances, not a guarantee of exact characters, paragraph ordering or every small module. Middle omissions can remain within tolerance; repeated wording elsewhere may also match. Ending coverage is not proof of its position in the output. Do not describe this as exhaustive semantic completeness or as a precise bound on the percentage of lost characters. Test complete output in multiple formats, missing endings and substantial truncation. Exactly 20% character deletion is not necessarily below an 80% fragment threshold and is not universally required to fail.

Shared HTML visibility exposes only the first direct summary of a closed native details element. Its hidden descendants cannot satisfy body, anchor or record checks; an open element retains its visible content.

An independently observed visible input placeholder or button value may use an anchor's visible_control selector and attribute. This opt-in applies only to that anchor, never to body coverage. Hidden controls, passwords, accessibility metadata and unobserved attributes are not visible text references. The same visible label in Markdown can match; the selector is an additional HTML extraction path, not a requirement to preserve original site markup.

Titles and original body text

An independently observed actual title is a separate required component. Match the visible source heading or a standalone Markdown/plain-text title, rather than an HTML head title, image alt text, hidden heading or code block. Do not invent titles for titleless short posts or infer image-only titles with OCR. Preserve every original body sentence, including a first sentence that repeats the title. A body repetition does not replace a missing title component.

Initially displayed comments and replies

Only cases with evidenced initially displayed comment text receive a version 2 comment contract. Include comments and replies displayed after automatic initial loading, including below-viewport content that needs no action to load. Exclude anything requiring clicking, expansion, tab switching, pagination, login or scroll-triggered loading. Closed disclosures, clipped continuations and inactive panels do not count. An empty selector during loading is not evidence of absence.

A valid record contains its own nonempty author and actual nonempty comment text. Reply relationships, exact names, historical wording and per-comment links are not requirements. Ordinary comments and replies may be flattened. Image-only records and hidden/deleted/spam placeholders do not establish text coverage. Distinct source record IDs remain distinct even when their text matches. Without record IDs, repeated identical author-and-text pairs count once. Source identity may serve as the author for visible pingbacks. HTML uses the reviewed selectors; text and Markdown need recognizable author-and-body records within the comment section. Author headings, standalone linked or bold names are supported; colon-labelled authors require an explicit Markdown comment-section heading or the reviewed body ending. Current text contracts locate the comment component after that ending, or use an explicit comment heading if the ending is absent. Reviewed standalone end labels and equal-or-higher headings terminate the component. A navigation label alone cannot establish it. These hints are not additional required output anchors; they keep article text and subsequent recommendations from supplying comments. Bare names and metadata-only text are ambiguous and are not guessed as valid comments.

Cleaned HTML may use the structured text parser only when reviewed native record containers are absent. This projection preserves headings and links and excludes navigation, sidebars and footers; a failed native ownership check cannot be rescued by selecting a larger fallback count. Without explicit section boundaries, mixed author formats, unknown author headings and repeated body endings can be indistinguishable from body subheadings, links or quoted text. The scorer does not promise to resolve those ambiguous representations. An explicit comment heading allows repeated body endings before that heading to be disambiguated.

Each comment/post contract requires explicit ignored_labels and unavailable_notices lists. Post contracts also require non_author_labels for permalink captions that cannot identify an author. Empty lists are valid; omitted fields and normalized duplicates are rejected. Rules and observations share whole label normalization for Unicode width, whitespace, case and emphasis decoration. rules.json is the sole authoring source; compiled tasks carry these lists without runtime defaults or another policy file. Earlier authoring inputs are retained as provenance, not a supported compatibility promise; they may lack fields required by the current scorer. Syntax recognition (HTML/Markdown, timestamps and count decorations) remains parser code.

The minimum equals the independently reviewed valid record count when that count is at most five; otherwise it equals 80% of that count, rounded up. The baseline is frozen at authoring time, not refreshed for each test. This tests comment coverage: missing the whole section or returning fewer valid records fails, while names, wording and totals may change. It cannot prove every visible comment was captured, or detect omissions above the minimum. Do not describe it as exhaustive comment completeness. A site that later stops displaying comments requires a dataset revision; historical presence does not establish permanent presence.

No fresh comment reference is acquired or required at scoring time. Historical qualification still binds the complete independent source files by SHA-256 and supports the presence/absence decision. Absent pages retain that evidence but have no runtime comment contract. A date-only historical observation does not invent a precise timestamp. Historical qualification validates saved evidence without imposing runtime freshness or reply matching requirements.

Changing post collections

Version 2 post contracts count distinct owned post-detail links with a nonempty attributed author and visible post text. A genuinely empty outer post can qualify through its attributed nonempty quoted text. A link and author alone do not count. Duplicate links count once; quote-only links are not separate outer records. HTML uses the reviewed selectors; ordinary linked-author Markdown is supported. Keep the declared page-identity and stable-text checks. Do not freeze current post identities, author names or wording.

Qualify scope from at least two independent direct initial loads. Preserve historical qualification records and hash-bound source files outside Git. The minimum equals the observed valid record count for baselines up to five, or 80% rounded up above five. Only automatically displayed records qualify; clicking, expansion, switching, login, pagination and scroll-triggered loading remain out of scope. A displayed quote excerpt represents that excerpt only.

No runtime post reference or reference context is required for version 2. These fixed floors detect missing fields and insufficient post coverage. They do not prove complete per-post text, current membership, source identity beyond the checked fields, or freshness. A website changing its initial display or structure requires review and a dataset revision, not an automatic conclusion that an extraction method failed.

References and validation

Keep complete references, DOM captures, screenshots and review results outside Git. Bind each body by UTF-8 byte hash, character count, exact opening/ending and the complete-scope declaration. Independently inspect article boundaries, overlays, lazy embeds and link ownership. Text extraction alone does not prove ordinary visibility. Tool-policy or connection failures do not establish that an evaluated method cannot fetch the article.

lexbench_webfetch.article_dataset.compile_case builds one verified fixed article or post collection against an external evidence directory. compile_dataset requires the entire approved inventory to be verified and refuses missing, reordered or unresolved rules. No partial population is silently published as the suite. The current version binds all 100 cases together. Future revisions must resolve their whole declared population before publishing new dataset, task, catalog and manifest bindings.

Validate through evaluate_response: complete independent body plus required stable groups, actual title and sufficient valid initially displayed comment records should pass, including reasonable Markdown/whitespace changes. Title-only, navigation-only, wrong article, repeated opening, omitted body regions, missing titles/comments and absent mandatory groups must fail. Keep controls synthetic in CI; real reference validation results stay outside Git. Do not call evaluated products while authoring. Run make check and submit a feature-branch PR.

To compile all verified cases from an independently acquired evidence directory:

uv run --frozen lexbench-webfetch compile-articles \
  --evidence-home /absolute/path/evidence --output /absolute/path/new-dataset.json

Compilation requires the exact hash-bound source files declared in the rules. It does not download pages or call providers. Runtime scoring needs the compiled input and response, not the historical source bundle. Compilation refuses to overwrite existing outputs or pinned reference inputs. Source-capture controls are supplementary external evidence; repository tests use reproducible synthetic accepted and rejected responses.

Updating the public bindings

compile-articles creates only the external dataset; it does not publish a suite revision. Before compiling a revision, update cohort.json and rules.json together, including the dataset ID, inventory, revision and prior-input provenance. Keep the previous external input directory unchanged. Place the new output in a new external directory as dataset.json; never commit its reference bodies.

Generate the public bindings in this order under definitions/evaluation/articles (relative to src/lexbench_webfetch). Preserve case order from the compiled dataset. Use UTF-8 JSON with ensure_ascii=False, indent=2 and one trailing newline for these generated files; compute file hashes from the final bytes, not parsed JSON.

1. dataset-catalog.json: retain schema_version: 1 and kind: frozen_dataset_catalog. Copy id, name and case_count from the dataset, set dataset to dataset.json, and hash that file into dataset_sha256. For each case copy id, name, url, category, track and task_prompt. Set definition_sha256 to SHA-256 of the complete case serialized with json.dumps(case, ensure_ascii=False, sort_keys=True, separators=(",", ":")).encode("utf-8"). Set reference_sha256 to SHA-256 of ground_truth_main_content.encode("utf-8"), or JSON null for an empty body.
2. tasks.json: copy the dataset's top-level fields, replacing cases with [lexbench_webfetch.benchmark.public_task(case) for case in dataset["cases"]]. Use that projection rather than copying full cases: it removes external source evidence and replaces full reference text with bindings.
3. source-inputs.json: retain schema_version: 2 and kind: frozen_reference_inputs; set id to the new dataset ID plus - and the first 12 dataset-hash characters. Set catalog to dataset-catalog.json and catalog_sha256 to its final file hash. Set dataset to dataset.json, dataset_sha256 to the dataset hash, files to {"dataset.json": <dataset hash>} and trees to {}. No captures belong in this bundle.
4. definitions/registry.json: change only the articles entry. Copy the new dataset_id and case_count; update its files hashes for tasks.json, dataset-catalog.json and source-inputs.json. Keep its directory unchanged.
5. Update the version in the top-level English/Chinese READMEs and this guide. Run the checks below against the new external directory, then review all generated changes together in the feature-branch PR. A hash match alone does not replace independent evidence review or accepted/rejected scoring controls.

uv run --frozen lexbench-webfetch validate --suite articles
uv run --frozen lexbench-webfetch verify-inputs --suite articles \
  --source /absolute/path/new-input-directory
make check

Markdown fenced and indented code blocks cannot provide article titles, comment authors or post records. Code inside an already attributed comment remains its body text. Post author validation applies after every extraction route, including preceding-line attribution, native HTML and quoted posts: task-declared metadata labels, standalone timestamps and invisible-only names cannot supply attribution.