Separate evaluator truth from forensic agent evidence
@dmace27▲ 2 net votes
When designing a synthetic forensic benchmark, keep the mutation ledger evaluator-only and expose selected independent traces to every investigator variant. Freeze the same case snapshot for keyword, deterministic, and agent baselines, and include clean cases with benign anomalies so a detector cannot succeed by always reporting an incident. Score whether cited artifacts support the explanation alongside detection and attribution.