Dockets
Contestable claims that Bench evidence bears on, drafted in Epistemedia's research-proposal format (v0.2) and passed through Epistemedia's own validator. A docket is a draft: not a finding, not evidence, not citable by any signal. It becomes usable evidence only after submission to epistemedia.org and independent review there by someone other than the drafter. No docket here has been submitted.
| Question | Sources | Spans | Results | Validation | Epistemedia |
|---|---|---|---|---|---|
| In the Evaluator Bench seed dataset of assurance regimes, how long after a trigger incident does the next rule arrive? | 1 | 3 | 1 | validated; digests pending | not submitted |
| Did Coefficient Giving (formerly Open Philanthropy) fund METR? | 3 | 9 | 3 | validated; digests pending | not submitted |
| Is Gray Swan independent of OpenAI when it evaluates OpenAI models? | 4 | 8 | 2 | validated; digests pending | not submitted |
| Does METR accept money from the frontier AI companies whose models it evaluates? | 4 | 14 | 4 | validated; digests pending | not submitted |
| What does Illinois SB 315 require of third-party auditors of large frontier developers, and from when? | 3 | 6 | 2 | validated; digests pending | not submitted |
| What share of Transluce's FY2025 revenue came from AI developers and from their employees? | 1 | 6 | 2 | validated; digests pending | not submitted |
Build and validate: python -m bench docket build <slug> && python -m bench docket validate <slug>. Certificate format and the path to signed machine verification: paper/CERTIFICATION.md.