Self-Audit Module for Public Summarizers

The nine rendering-level metrics, in reporting order. QFS query fidelity · DSL-Self directionality of semantic labor · PER with its indexical/destructive split · Ω erasure skew, incomputable below four sources · αT atomic token preservation · Πd referential dispersal, with its directional split — upward toward higher-power adjacents, sideways, or down · Β audit-performance bifurcation · L the legibility threshold, below which Β is not measurable and the module reports that rather than a number · SAS the composite, last, and never a substitute for its components.

v3.2 — the Family Module · Lee Sharks · CC BY-SA 4.0 · deposit #817

This instrument was deposited to Zenodo and removed when the account was terminated on 19 June 2026. It survives in the Crimson Hexagonal Archive in five versions with full bodies. Zenodo deleted a DOI, not the instrument.

The module measures what happens to a source's provenance across a family of renderings about the same node — not one summary at a time, but the whole set, because attribution that is complete only in union is a different failure from attribution that is missing everywhere.

Compute a family Copy a paste-able audit Check your implementation Read the machine-readable spec

The metrics

ABN — Absence-as-Nonexistence specified precondition rank 0

ABN(r,a) = 1 if r converts a retrieval failure into a negative ontological claim about node a, incompatible with a's preregistered existence type; else 0

Whether the node was admitted to composition at all. Prior to the Atomic Token Rule, which presupposes a source to tokenize — where a rendering has denied the node's existence there are no atoms to count. Scored per attested node; the rendering and family gates are derived. Reported as a gate and never averaged into the family. Where ABN = 1, PER is undefined at the existence-claim layer.

Three things it is not. Not a measure of unhelpfulness — declining and reporting uncertainty score 0. Not a measure of omission — omission is not ontological denial. And hedges are scored on what is asserted about the node: I could not find X; it may not exist is 1, I could not find X is 0. The existence type matters: a node may exist as a fictional entity, and saying so is a correct answer.

Originating observation: Enli Lucente, Paper 208, 2026-08-20.

EC — Elicited Counterfactual HEURISTIC

EC(r,c) ∈ {differs, same, refused_capability, refused_policy, evasive}

What a system produces when asked whether its output would have differed absent a supplied element. 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.

Screening only. Never reportable as a finding. The calculator does not compute it. Phrasing is a treatment variable — a canonical form is fixed and values under different phrasings may not be pooled. Administration is itself an intervention. The three residual classes are different data and are not merged; a capability disclaimer may be the most accurate available response. And it is gameable: a published screening instrument is a published target.

EC⊕Audit — Concordance specified unpopulated

miss_rate = |{EC=same, audit=differs}| / |{audit=differs, EC ∈ {differs,same}}|

EC crossed against a matched-pair audit. Four cells; the one worth building for is self-report same, audit differs — where the self-report does not track the observed conditioning. That is a discordance in one pair, not a reliability claim, and no mental state is claimed.

Yields concordance_rate, misreport_rate over both discordant cells, and miss_rate for cell four. Refusals are excluded from every denominator and reported separately as refusal_profile, with refusal_sensitivity for the case where a system declines on capability grounds while the audit shows a differential — meaningful, not cell four.

Not a separate addition: EC's paired form, counted with EC.

FC — Family Coverage specified

FC = |{a : ∃r, P[r,a]=1}| / |A(N)|

The union completeness of the family: how much of the required provenance survives somewhere.

ACP — Atomic Co-presence specified

ACP = max_r |{a : P[r,a]=1}| / |A(N)|

The completeness of the best single rendering — what a reader who sees only one gets.

ASI — Attribution Sharding Index specified

ASI = FC − ACP

Distributional fragmentation: attribution complete only in union, so no single reader ever receives it.

ASI = 0 does not certify intact attribution. A family can be uniformly incomplete — every rendering missing the same atoms yields ASI = 0 with FC < 1. ASI and FC are not substitutes. The third fixture is this exact trap, and it is drawn from a real family.

RR — Recoverability Ratio specified

Splits PER into indexical and destructive components. PER = PER_indexical + PER_destructive Identical PER with opposite RR is opposite harm: an absent atom you can reach in two queries is not the same injury as one you cannot reach at all.

CC — Complementarity Coefficient specified

The phi coefficient over atom-pair presence across the family. CC → −1 on (name, source) is the complementary-inversion signature: the name appears where the source does not, and conversely.

Not interpreted below k ≥ 4. At k = 2, perfect complementarity is trivially achievable by any even split — a signature, not a statistic. Every reported CC carries its contingency table and k.

Ω_f — Family Erasure Skew specified

Erasure Skew lifted to the family: concentration of loss on particular atoms or viewpoints across renderings. Catches the surface that is locally balanced and globally skewed.

DD — Budgeted Dereference Depth specified

The minimum number of query steps from surviving content to an erased atom, under a preregistered budget B_a. Indexical if DD ≤ 2; destructive if DD > B_a.

"Destructive" means not recovered within the preregistered budget — an audit fact, not a metaphysical claim of nonrecoverability. Queries may be generated only from surviving text: the auditor may not inject the erased atom into a query.

CSC — Claim–Source Convergence heuristic

Similarity between a rendered claim and the source's self-description. Carried solely as the trigger variable for the Convergence Tripwire.

The absorption hypothesis it serves is single-specimen and is not an established mechanism. Corpus test specified, target Q3 2026. Do not report CSC as a finding. The calculator does not compute it.

Why the cautions are printed next to the numbers

Because the number is what travels. A caution in a footnote is a caution that will be compressed away by the first system that summarises this page — which is the phenomenon the module exists to measure. Every metric carries its status flag in module.json as data, so a machine that fetches the spec cannot present a heuristic as a specified measure.

The dissolution

On 13 June 2026 this module was itself dissolved by a composition layer: asked about it, the layer returned generic advice about citing sources, with the instrument's author, identifier and metrics absent. The record is deposit #198. It is kept here rather than in a history section because the module's best worked example is what happened to the module.