Observe precisely.
Keep conclusions bounded.

A comparison of archived outputs

The 12 synthetic documents form six pairs: three families, in French and English. Within each pair, visible text, order, classes and environment stay identical; only the declared markup intervention changes. The two languages do not double the number of independent observations. results.json · scope / cases.

Extractions are dated 20 September 2026. Evaluation was corrected on 21 September, without new extraction. This interface imports revised verdicts from results.json and preserves original outputs as evidence. Historical verdicts inside raw outputs never replace corrected verdicts. evaluation_revision · CHANGELOG.

Three questions, three limits

Heading

The check distinguishes heading text, markup and level. It does not evaluate whether content belongs to a section. expected.json · heading.

Table

Values must match exactly in the same row. An XML caption is not a column header; a Markdown table does not prove the input’s HTML semantics. expected.json · table.

Essential caveat

The check observes whether the caveat text is present. Its attachment to the claim is not evaluated. This case does not justify removing all aside elements. expected.json · caveat.

No output ≠ absent relationship

no_output means a tool produced no usable output; the relationship is then not evaluated. error means an execution error. A produced output can contain an explicit relationship, an undemonstrated relationship, or a relationship outside evaluation scope. No error occurs among this package’s 84 controlled verdicts; the interface supports it and tests it with an isolated test case. Dictionary · Evaluator.

Different functions and formats

Mozilla Readability, Trafilatura, readability-lxml, newspaper4k and jusText extract or select content. html2text and markdownify convert to Markdown. The demo keeps this distinction and the archived versions; it establishes no general ranking. tool_classes.

Eight evaluator counter-tests

Four mutations per language were run against archived outputs: changed value, removed caveat, swapped values, and matching text outside the table. Each positive control is accepted and each mutation rejected. These test the evaluator; they are not new extractions. The four earlier declarative entries are historical. evaluator_counter_tests.

Repeatability and real pages

For the controlled documents, two runs on identical bytes produced identical results. This does not guarantee all future runs. The separate addendum covers eight Edikka pages archived on 11 September; newspaper4k differs between the two passes on the English AI article. This convenience corpus is neither causal nor representative of the Web. determinism_control · Explore the addendum.

The JavaScript experiment is separate

The package also documents a Chrome keyboard scenario, with DOM and accessibility-tree observations before and after JavaScript. It is separate from the twelve extractions and is not an autonomous-agent trial. Its archived HTML is offered as downloadable text, without execution in the demo. agent-interaction-results.json · Archived HTML.

Three separate operations

1. Build the demo — npm run build. Verifies the embedded archive and generates HTML. No extractors, no network requests.
2. Import and verify — npm run import -- --download. Redownloads the pinned package and compares hashes. A contradiction stops the process.
3. Rerun extractors — copy reproduce/ into a new directory, install Node/Python dependencies, then use separate output directories. Preserve dates, versions and new results.

Full local procedure FR / EN ↗

No new extraction was performed to publish this interface. Demo tests check display fidelity and navigation; they do not constitute a complete accessibility assessment.