Audit a Coval evaluation

Inspect what the customer actually ran and what decision those results support. Stay read-only unless they separately request a concrete repair or test.

Collect the smallest useful evidence

Confirm organization, workspace, API environment, agent, time window and decision. Use explicit resource IDs where available. Start with a recent relevant run, its agent, persona, cases and metrics; sample completed, failed and unscored conversations. Don't download the whole account or launch a run to fill a gap.

Discover installed commands with coval --agent agent manifest and resource context. Common reads: