Agent Eval Platform · evidence explorer
Recorded attempts, with the checks behind each verdict.
Coding agents, a pull-request reviewer and fixed controls, each graded by checks it could not see or change. Every square below is one attempt. Select any of them to read its checks, timings and receipt chain.
All recorded attempts
Loading evidence
Collections
Loading
Why attempts were not accepted
Trial matrix
One cell per first attempt. Select a cell to open its receipt.
Timing
Replay a decision
Add a resource limit and watch the verdicts change.
The preview reuses saved measurements, so nothing reruns. Behavior checks stay required, and the recorded verdicts never change.
seconds
A limit needs an independent measurement. Attempts without one become inconclusive in the preview.
Investigate
Every attempt in this collection
| Task | Policy | Outcome | Time |
|---|
No attempts match these filters.
Select a row for its receipt
How each attempt is graded
See the architectureReserve the trial
Task, policy, recipe, budgets and evaluator identities are frozen before anything is dispatched.
Observe independently
Behavior and final state are checked outside the candidate, which never sees the expected answers.
Keep the denominator
Failures and missing results stay in every count. Assisted corrections are reported separately.