pub struct CandidateEvaluation {
pub candidate_id: CandidateId,
pub dataset_version: String,
pub replay: ReplaySummary,
pub baseline_report: Value,
pub candidate_report: Value,
pub verdict: EvaluationVerdict,
pub thresholds: EvaluationThresholds,
pub evaluated_by: ProvenanceAuthor,
pub evaluated_at: DateTime<Utc>,
}Expand description
The journaled evaluation payload — the CandidateEvaluated event’s
output. The evaluation is evidence, not a log line: the verdict, the
report pair, the dataset version, and the replay fixture ids, so the
improvement is explainable afterward by walking ids (candidate →
evaluation → reports → fixtures).
Fields§
§candidate_id: CandidateIdThe evaluated candidate.
dataset_version: StringThe versioned dataset both reports ran against. Versioned, never edited in place — the envelope’s auto-promotion bar may name the exact version a clean verdict must come from.
replay: ReplaySummaryThe replay half: divergence summary plus the fixture ids.
baseline_report: ValueThe baseline experiment report, verbatim (rusty-eval’s
ExperimentReport serialized; carried as a value because the
runtime does not link the eval crate — see the module docs).
candidate_report: ValueThe candidate experiment report, same convention.
verdict: EvaluationVerdictThe comparison verdict the gate reads.
thresholds: EvaluationThresholdsThe thresholds the comparison applied, echoed so an auditor reads the bar the verdict was judged against, not a later default.
evaluated_by: ProvenanceAuthorWho ran the evaluation (attribution is mandatory here too).
evaluated_at: DateTime<Utc>When the evaluation ran.