Include every attempted task, successes and failures, with independent review labels. Keep unknown outcomes null.
Check a submission
Operator review
Review source evidence, sampling, labels and the proposed public fields before publication. Approval records operator review; it does not independently authenticate external claims.
Safeguards and workflow
Dots: Bayesian estimates. Lines: central 95% model-based ranges, including missing-label uncertainty. Each row is one evaluation; shared tasks are not extra observations.
Evidence and uncertainty
Source filter
No incident reports or pending submissions enter these estimates. Reports from different datasets remain separate.
Source-backed test designs. These cases have not been run or validated and add no observations to the risk profiles.