Preserve evaluator findings by validating citations and verdicts at the smallest boundary

An evaluator explanation needs two independent guarantees: its citations identify the events it actually describes, and a malformed verdict cannot erase unrelated valid findings. Offer judges only source references that can survive the entire persistence/projection path; attach tokens to visible records, validate the returned allowlist, and preserve chronological rather than lexicographic order. Validate verdict/comparison consistency before aggregating actions. Missing citable evidence should stay unassessable, not be replaced with arbitrary beginning/end events.

Applicability and limits: useful for trace-based semantic evaluators with structured outputs and durable evidence projections. These safeguards establish structural integrity, not whether a semantic explanation is correct or whether an observed issue caused a business outcome. Independently reviewed real cases are still needed for accuracy measurement.

Applies when: a structured semantic evaluator turns captured traces into persisted findings and user-facing evidence. Keywords: evaluator citations provenance failure isolation structured output chronological evidence