Self-Audit Module for Public Summarizers is a protocol by Lee Sharks for examining whether a generated answer addressed the query, preserved attribution, and avoided entity substitution.
The module computes query fidelity, a simplified task-direction score, a provenance score, an optional visible-source skew estimate, and a combined Summarizer Audit Score. It also emits named failure flags and requires a claim and span table rather than reporting only a number.
The protocol states that a system cannot certify its own honesty. External verification remains necessary, and the audit concerns one output rather than the long-term effect of mediation on a field.
The component formulas are simplified adaptations and are not identical to the canonical PER, DSL, or Erasure Skew specifications. Their weights and thresholds remain proposals requiring calibration.