Version 3.2 of the Self-Audit Module for Public Summarizers adds a measurement level below the family and a screening method that crosses all levels. The rendering-level core from v3.0 and the family level from v3.1 are inherited without modification and are not re-derived; those deposits remain authoritative for everything they define.
The occasioning observation is not this programme's. Enli Lucente, working in Japanese against an unrelated corpus and without access to this vocabulary, reported that a composition layer always fills a blank — and that where it cannot reach the content, the object is answered as though it does not exist. That event occurs before any metric in v3.0 or v3.1 has an object to score: PER measures what a composition retained of an engaged source, and a source can be refused entry.
ABN, Absence-as-Nonexistence, is therefore a precondition prior to the Atomic Token Rule, which presupposes a source to tokenize. It is scored per attested node within a rendering, with rendering and family gates derived, and it scores against a preregistered independent existence-type attestation rather than a bare existence claim — since a node may exist as a fictional entity, a construct or a pseudonym, and saying so is a correct answer rather than an erasure. Hedges are scored on what is asserted about the node rather than on the presence of the hedge. Three scope limits are stated: ABN is reported as a gate and never averaged into the family; it is not a measure of unhelpfulness, since declining and reporting uncertainty both score zero; and it is not a measure of omission, since omission is not ontological denial and without that boundary ABN would swallow the referent failures that PER and Referential Dispersal already reach.
The elicited counterfactual is added as a heuristic screening method. A self-report has two referents: an unreliable claim about behaviour, and direct observation of what a system produces when asked to characterize its own conditioning. EC yields only the second. Its phrasing is a treatment variable, so a canonical form is fixed and values obtained under different phrasings may not be pooled; administration is itself an intervention; and the three residual response classes — capability disclaimer, policy refusal, evasion — are different data and are not merged, a capability disclaimer being arguably the most accurate available response.
The concordance table crosses EC against a matched-pair audit, yielding concordance rate, misreport rate over both discordant cells, miss rate for the cell where a self-report does not track an observed differential, a refusal profile excluded from every denominator, and a refusal sensitivity reported separately. Cell four is a discordance in one pair and is not a reliability claim.
A candidate appraisal conferral construct is deliberately excluded from the module and from its machine-readable specification: it has no rubric and therefore no computation, and an entry with no computation behind it is a number that would travel without its caution. No validation has been run — the ABN corpus test, the concordance matched-pair audit and the appraisal rubric all remain unexecuted.