Persona Fidelity Audit

Audit how well agents' actual output matches their persona text. Pairs the mechanical scorer (fingerprint + embedding similarity, populated on every assistant message at insert time) with the LLM-as-judge rubric for spot-checks on outliers. Produces an evidence-based report: which agents have flattened into generic LLM telemetry, which schemas are doing the flattening, and what to fix.

Details

Audit how well agents' actual output matches their persona text. Pairs the mechanical scorer (fingerprint + embedding similarity, populated on every assistant message at insert time) with the LLM-as-judge rubric for spot-checks on outliers. Produces an evidence-based report: which agents have flattened into generic LLM telemetry, which schemas are doing the flattening, and what to fix.