Eval candidates

An eval candidate is a proposed behavior definition derived from one or several verified failure groups. It is not an executed test result.

• one or more representative cases;
• proposed input or fixture description;
• expected behavior;
• rubric and grader proposal;
• source failure groups and findings;
• evidence spans and telemetry gaps;
• immutable generator and detector provenance;
• explicit review state.

From Investigation, choose Create eval candidate. Review the evidence packet, expectations, rubric, grader, source run/revision, and telemetry-gap summary before creating one draft.

From Failure Inbox, select several exact groups and choose Create evals. The preview reports:

• proposed candidates;
• existing candidates that will not be duplicated;
• excluded groups and reasons;
• representative occurrence chosen per group.

Create the batch only after reviewing the preview. You do not need to select every trace individually.

Perseval selects a bounded representative occurrence using deterministic group evidence. Exact groups remain canonical even when feature similarity is enabled. Candidate identity and evidence hashes do not change because a secondary cohort was rebuilt.

Drafts enter the review queue as Pending. A reviewer can:

• Accept a definition for downstream use;
• Reject it with an explicit decision;
• Defer it while keeping it in the queue.

Decisions and reviewer reference persist across restart. Review does not activate a grader or mutate the source finding.

Generation is idempotent for the same source identity. Open the queue and find the existing definition instead of creating a duplicate.

Choose one explicit project, confirm a concrete evidence occurrence exists, and resolve severe telemetry gaps before retrying.

Check project scope and queue-state filters. Accepted, rejected, deferred, and pending records are durable.