Failure analysis

Perseval starts from structured behavior, not from “all red spans are defects.” The goal is a defensible claim connected to exact evidence.

• An error span is producer telemetry: an operation reported Error.
• A finding is a detector claim supported by bounded evidence and a stated certainty.
• A failure group collects findings with the same exact deterministic signature.

An error can remain outside the Failure Inbox when work was optional, recovery was verified, or required behavioral facts are missing.

Detectors consider operation identity, requiredness, status/result, effect, retry safety, attempts, observed state, root outcome, events, and links as relevant to the rule. They do not infer these facts from free-form assistant text or a tool name.

The Finding inspector reports:

• detector and version;
• severity and recovery;
• exact, bounded-inference, or inconclusive rule match;
• missing facts and semantic coverage;
• evidence references and provenance.

Examples include a required operation that ended without compatible recovery, an uncertain mutation that was not verified, repeated tool failure, or a loop. The precise detector set can evolve; detector version stays attached to immutable analysis identity.

Recovery is evidence, not optimism. A later successful call must be compatible with the failed operation and state. Compensation or verification should be explicit. A root “completed” claim does not erase an unrecovered required failure.

Exact failure signatures remain canonical and drive recurrence, review, and eval candidates. Optional feature-similarity cohorts are secondary navigation metadata; they cannot merge exact groups or change draft eval identity.

Recurrence uses a shared eligible-run denominator for the active project/environment/build/session/time scope. It answers “in how many eligible runs did this exact pattern occur?”

Trends compare rates across observed time buckets, not raw counts. No eligible population is different from a true 0% failure rate. Recurrence measures repetition; it does not prove that a diagnosis is true.

If an errored run produces no group:

1. open it from Runs;
2. expand the relevant branch in Full Trace;
3. select the error span;
4. read the missing telemetry named in the Finding inspector;
5. inspect Attributes, instrument the missing facts, and send a new run.

Do not rewrite old finalized evidence. The new run provides a reviewable comparison.