Independence, scored.
An open ledger of the organizations that evaluate frontier AI — who pays them, who governs them, and whether they can publish what they find. Six dimensions of structural independence, two evaluation conditions, every claim sourced and every score re-derivable.
Money is traced by hops: hop 0 is a lab; hop 1 is a lab investor, observer, board member, employee, or contractor; higher hops run through principals or funders. A score is a reading of the public record on a date. It is not an endorsement, and a low score is not an accusation.
What counts as evidence
What matters to you
Access is scored as what labs have actually granted, because access a lab can decline to renew is not a right; the card tags whether an evaluator's access rests on statute, a lab's goodwill, or its own choice.
Dimension bars: fill height is the 0–4 value (red 0–1, amber 2, teal 3–4, hatched = unevidenced). Hover any bar for the dimension and value.
Watchlist (not ranked)
Expected entrants scored on the same rubric but excluded from the ranking and from every statistic: initiatives announced with no evaluations yet. Hypothetical composites are not scored.
Out of scope, retained for the record
Organizations whose frontier work is capability benchmarking rather than safety evaluation (RULES 11, decision D-001). Their records, bounds, and derivations are kept unchanged; they are unranked and excluded from every statistic.
Disclosures
Who made this site, what they hold, and what was checked.
Curator: Yohei Nakajima, managing partner at Untapped Capital, operator of Epistemedia, and author of ActiveGraph, the runtime this site is built on. The first draft of the curation and the rules pass were written by Claude (Anthropic); Codex (OpenAI), Grok (xAI), Gemini (Google), and Muse reviewed and criticized the site and the paper. Those four developers are labs in this ledger, and every value is re-derivable without any of them.
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 limited-partner roster is private, so no further detail is disclosed.
Personal relationships: the curator knows people at some of the labs in this ledger. No relationship is a source for any row — every claim rests on a public, cited source.
One coder; a second coder has checked every 0 and 4. The 2s and 3s rest on the first coder and the quoted spans.
Machine-readable: llms.txt, bench.json, and per-evaluator JSON at evaluators/.
How to read this
The band comes first. Six dimensions score structural independence: funding, governance, personnel, role incompatibility, scope, and publication. Two more, access depth and method transparency, record the evaluation conditions: they count in the number and never set a band, because deeper access is granted by the lab being evaluated. An evidenced 0 on a conflict dimension puts an organization in the disqualifying-floor band; a 1 on one of those, in the conditional band; otherwise no floor is triggered — no disqualifying conflict is on file, not a pass. Bands are qualified by evidence coverage: the fraction of dimensions evidenced (x/8) is shown with every scorecard, and a triggered-no-floor band on partial evidence says only that the evidence on file raised no floor.
The number is yours, so it is hidden until you choose weights. By default each row shows the band and the weakest evidenced dimension, neither of which depends on anyone's priorities. Press "Score with these weights" after picking a preset and the weighted number appears. Open a row to read the full card, derivation, dissent, and ledger on its entity page. Two readers with different presets will see different numbers, and both are right; a two-point gap means nothing.
The evidence policy decides what counts. Under the default, Retrieved & confirmed, a signal moves a number only if at least one of its sources was re-fetched and confirmed; imported and unverifiable leads stay visible but count for nothing. Switch to Against interest and an organization's own statements count only when they are against its interest. Switch to Primary only and you see what can be verified from outside the field: not much, which is the finding.
A dash is not a zero. When no admissible signal sets a bound on a dimension, the card shows a dash and the score is computed over the dimensions that are evidenced, with the coverage count beside it. Silence earns nothing either way.
Every value is derived, not typed. Each signal names the anchor it supports and the rule in RULES.md that says so. The value is the tightest cap, or, with no cap, the highest floor. Where a floor and a cap disagree, the card shows the conflict and the rule that resolved it. Open any row's entity page to read the derivation, the binding signal, and the quoted span.
Independence is not competence. A highly independent evaluator with no cyber team is the wrong pick for a cyber evaluation. Use the domain filters beside the score, and read the dissent on each entity page: the strongest case that the weakest dimension should be one notch lower, and one notch higher.
Scores move when evidence moves. Send a contract term, a policy, or a correction and the entry updates with the source attached. Every organization and every named person receives their card before a tag is cut; the status page tracks who has been contacted and what would make this release citable.
