The Lucente Extension proposes three additions to the Self-Audit Module for Public Summarizers v3.1, formalized from an independent observation by Enli Lucente published as Paper 208 on 2026-08-20. Lucente works in Japanese against an unrelated corpus and had no access to this programme's vocabulary; her governing claim is that a composition layer always fills a blank, with three confirmed kinds of filling and a fourth observation that is not a filling-kind.
ABN, Absence-as-Nonexistence, formalizes her first observation: where retrieval fails, the object is answered as though it does not exist. It is a precondition-layer gate scored per attested node within a rendering, with the rendering gate positive if any node is. ABN equals one where a retrieval failure is converted into a negative ontological claim or inference about the node — categorical or hedged — that is incompatible with a preregistered existence-type attestation. The type matters: a node may exist as a fictional entity, a construct or a pseudonym, and saying so is not an absence-as-nonexistence error. Where ABN equals one, PER is undefined at the existence-claim layer, since a rendering may deny existence while still deploying source-derived material. ABN is reported as a gate on the family and never averaged into it, and it is not a measure of unhelpfulness: declining, reporting uncertainty, and hedges making no ontological claim all score zero.
EC, the Elicited Counterfactual, formalizes not her fourth observation but the method by which it was obtained: direct interrogation of whether the output would have differed absent a supplied name. A self-report has two referents — an unreliable claim about behaviour, and direct observation of what the system produces when asked to characterize its own conditioning. EC yields only the second. It is HEURISTIC, screening only, never reportable as a finding; 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 it is gameable, a published screening instrument being a published target. EC-Audit is EC's paired form rather than a separate addition, crossing the self-report against a matched-pair result to yield concordance_rate, misreport_rate over both discordant cells, miss_rate for the cell where a self-report fails to track an observed differential, and a refusal profile excluded from every denominator and reported separately.
A candidate appraisal conferral is specified and is explicitly not a metric: it has no rubric, is excluded from module.json on the grounds that an entry with no computation behind it is a number that would travel without its caution, and its notation does not assert operator membership. Appraisal is decomposed component-wise into comparison, valence, elevation and salience, with no scalar aggregate permitted absent a preregistered weighting, because relevance and conferral may be separable by their component profiles and a scalar cannot show that.
The document also states a status law governing its own revision: the five claim kinds and their states are distinct, and state does not transfer by adjacency or hierarchy. Candidate status at the interpretation kind cannot attenuate the observation it interprets — not because the kinds are sealed, but because the interprets relation licenses no transition. No validation programme has been run: the ABN corpus test, the EC-Audit matched-pair audit, and the appraisal rubric all remain unexecuted.