# Evaluator Bench A curator's ledger of third-party AI-evaluator independence: six structural-independence dimensions (funding, governance, personnel, role incompatibility, scope, publication) plus two evaluation-condition dimensions (access depth, method transparency). Each evaluator gets 0-4 per dimension, derived as the tightest admissible cap (or highest admissible floor) from bounded public signals under written rules — re-derivable by anyone from the quoted spans. An entry is not an endorsement; a low score is not an accusation; a 'no floor triggered' band means no disqualifying conflict is on file, not a pass. Curator: Yohei Nakajima (Untapped Capital). ## Machine data - bench.json — everything: 27 evaluators, 370 signals, 251 sources, rules, evidence policies, weight presets (0-4 ordinal, not ratios). - evaluators/index.json — slug, name, type, status, band, and JSON URL for every evaluator. - evaluators/.json — one evaluator's full record: assessments with per-policy derivations, binding signals with verbatim quoted spans (<120 chars), and the sources those signals cite. - exposure.json, timeline.json — ledger exposure and the regime timeline. Slugs match the entity pages: evaluators/.json <-> entity/.html. ## How to read it Method: /method/. A 0 on a conflict dimension needs a quoted span; an extreme (0 or 4) is second-coded. Values are derived, never typed; the verifier rejects stored values that disagree with their derivation. Scores move when evidence moves: open a PR with a source. ## Disclosure Holdings: publicly disclosed shares in Google, Meta, and SpaceX (owner of xAI). Untapped Capital's funders include holders of shares in private AI labs; the LP roster is private, so no further detail is disclosed. The curator knows people at some of the labs in this ledger; no relationship is a source for any row. Full: DISCLOSURE.md. Built 2026-09-15T00:00:00Z (RULES.md v0.1 (2026-09-15)); 24 of 27 evaluators ranked.