Evidence that can travel with the claim.

Use the strongest label the evidence supports, never the label that sounds most persuasive.

Name what was measured, when, where, with which denominator and what the measurement cannot establish.

A bounded case can reveal a mechanism worth testing. It does not become a universal benchmark by being precise.

Keep an unavailable value unavailable. Do not replace an evidence gap with a convenient estimate unless the assumptions are explicit.

Pricing, cache behavior, model names and provider rules drift. Date every time-sensitive claim and re-check it before promotion.

Any number used in public promotion must carry enough context for a reader to challenge it.

Ask for the artifact, validation method, missing field and the result that would change the conclusion. A claim that cannot survive those questions is not ready for promotion.

When a material claim changes, retain the previous wording, the reason, the evidence date and the new limitation.