How CompareLLM updates every day

An administrator records model facts and benchmark observations through the Catalog and Evidence desks. AI suggestions remain proposals and never write rankings without approval.

If nothing material moved, the site stays as it is. If an administrator approves a material price, score, speed, or new-model change, we write a changelog row and revalidate the affected pages.

Commercial facts and default-effort speed can be fetched from OpenRouter on an explicit refresh, and arrive as proposals. An administrator approves or rejects each one; a marketplace feed never displaces a better-sourced value, and it never touches identity, release dates, aliases or benchmark results.

Benchmark results are filed separately on the Evidence desk with their publisher, exact variant and version, evidence tier, and — for a first-party run — its result package. An unmatched OpenRouter id becomes a Catalog candidate, never a new model, so one marketplace listing cannot create forty thin compare pages overnight.

Publishing /news posts. AI may draft one after an administrator approves that generation, but the draft is private and a separate publish action makes it live. Approving the generation is not publishing it.

Every rating, too, in the sense that matters: the recipe behind each category Elo is authored and versioned by a person, and changing an input or a weight requires a cited research note.