Every factual page is built from an indexed research corpus. The corpus preserves source URLs, dates, evidence grades, disagreements, and fields that require periodic rechecking.
Evidence hierarchy
- A — First-party evidence. Official documentation, changelogs, API references, model repositories, licenses, and research papers from the model publisher.
- B — Reproducible evidence. Tests with a disclosed model version, settings, prompt set, harness, and date.
- C — Independent reporting. Credible secondary sources that provide useful context but are not the source of a product specification.
- D — Community signal. Reddit threads, forum posts, and individual usage reports. These can reveal pain points and test ideas, but do not become specifications.
- R — Rumor. Unverified release claims and screenshots remain labeled, quarantined, or excluded.
Dynamic claims and corrections
Prices, model IDs, context limits, release states, and supported modalities can change. We show a last-verified date and recheck affected pages after an official change. Material corrections change the page rather than silently preserving a known error.
Vendor benchmark scores are labeled as vendor-reported. A score is not described as independently verified unless a comparable third party has reproduced it with enough detail to evaluate the run.
Advertising boundary
Advertisements are labeled and visually separated from editorial text. Ad placement does not influence conclusions, evidence grades, comparison results, or inclusion in an article.