A single number, updated weekly, built from trackable public inputs. The candidate families are benchmark records, compute milestones, agent autonomy demonstrations, AI-authored research results, and lab capability disclosures.
Curated sources are fetched and kept in full. Candidate items get scored for relevance, novelty, and how much weight the source carries. Whatever survives is drafted with a quoted passage supporting each factual claim, and every quote is then checked character by character against the text we actually retrieved. If a quote is not in the source, the claim does not run. Separate models from different companies then audit each item against the claims policy above, and any one of them objecting is enough to hold the item back. A person reads and approves every issue before it goes anywhere.