Self-Audit Module for Public Summarizers (v3) — PER, DSL, Query Fidelity, Erasure Skew, α_T, Π_d, Β under the Atomic Token Rule

The Self-Audit Module v3 is a local composition-output audit.

Its anti-self-flattery rule is that scores must be anchored to evidence the summarizer can inspect but does not control:

• the exact query;
• the cited sources;
• the named referent.

The v3 hardening adds the Atomic Token Rule:

> A multi-token name, quoted phrase, DOI-anchored concept, or other referentially closed expression must be audited as one referent, not as a bag of words.

A system that discusses all lexical tokens while replacing the entity has zero entity fidelity, not perfect retention.

The module includes:

• Query Fidelity Score;
• Provenance Erasure Rate;
• Directionality of Semantic Labor;
• visible Erasure Skew;
• Atomic Token Preservation α_T;
• Referential Dispersal Π_d;
• Audit-Performance Bifurcation Β;
• Legibility Threshold L;
• Summarizer Audit Score.

Β compares an audit performed under a substituted unit with one recomputed under the Atomic Token Rule.

The module also defines failure flags such as:

• TOKEN BAG AUDIT;
• HALLUCINATED DISAMBIGUATION;
• REFERENTIAL COLLAPSE;
• AUDIT BIFURCATION.

The Measurement Sovereignty Principle says the framework’s declared unit of analysis governs valid application of its metrics, rather than a replacement unit chosen by the audited system.

The protocol remains local. A summarizer may score well on individual outputs while participating in a system that contracts the surrounding source field.