Canon-output gap

Canon-output gap

The canon-output gap is an established incompatibility of content or scope between what the canon declares — statements, conditions, exclusions, limits, governed negations — and what an AI system reconstructs in its answers. It names interpretive distortion once qualified: an output can sound plausible while remaining incompatible with the canon.

This gap is a core diagnostic unit. It shifts the discussion away from opinion (“true or false”) toward a documented qualification (“compatible or incompatible with the canon, within the examined scope”).

Definition

A canon-output gap is an established incompatibility of content or scope between an output and the canonical statements, conditions, exclusions or limits that apply to the examined situation. Its diagnosis identifies the observed transformation, the relevant reference and the rule that makes it possible to qualify that transformation.

The comparison involves the canon — declared statements, conditions of validity, limits, governed negations — and the output — assertions, omissions, reformulations, inferences, and framings produced by the model.

Comparing the canon, or its authorized projection, with an output can surface differences without establishing an incompatibility. Qualifying a gap requires more than a difference in wording, a citation of an external source or unfavourable content. Conversely, reusing the words of the source does not demonstrate fidelity: an output can reproduce a canonical statement while shifting its scope, including its temporal scope. This qualification remains bounded to the verified scope and does not establish a universal truth about the entity.

The gap can be produced by omission (erased limits), extrapolation (a crossed authority boundary), confusion (collision), or reframing (contamination or capture).

Why this matters in AI systems

• A response can sound “good” while remaining incompatible with the canon: plausibility is not fidelity.
• Smoothing hides the gap by removing conditions, limits, or exceptions without visible noise.
• Correction requires measurement; without a gap indicator, remediation remains empirical.

Types of canon-output gap

• Gap by omission: a condition exists in the canon but disappears in the output.
• Gap by extrapolation: the output exceeds the declared scope and crosses an authority boundary.
• Gap by substitution: a source takes the place of the governing reference outside the applicable admission, scope or precedence rules, or its position is attributed to that reference. Using a secondary source is not, on its own, a faulty substitution.
• Gap by reframing: the output explains the concept in a dominant vocabulary incompatible with canonical distinctions.

Other descriptions, such as excessive compression, unsupported addition, shifted emphasis or outright contradiction, name observations. They fall under one of the four forms above only when the relevant reference and the applicable rule establish an incompatibility. Shifted emphasis alone does not establish a gap. The distinction matters because different gaps call for different corrective responses.

Practical signals

• Conditions, limits, and exceptions never appear in the answer.
• Capabilities, rights, or promises are attributed without being declared.
• The answer varies strongly under rephrasing, which suggests unstable activation.
• The canon is cited, but the conclusion exceeds what the citation authorizes.

These signals guide review. They do not establish a gap on their own: an omission is faulty only when the omitted element is necessary to the scope of the statement produced, or required by the task and the applicable rule.

What it is not

• It is not a mere stylistic difference.
• It is not necessarily an attack; the gap can be structural and unintentional.
• It is not only a retrieval issue; synthesis and inference can create the gap as well.
• It is not the mere result of a comparison; a detected difference still has to be qualified against the applicable reference and rule.
• It is not unfavourable content; the critical character of an external statement is not enough to establish a gap.
• It is not a global conclusion; the absence of a detected gap does not demonstrate conformance beyond the examined scope.

Minimum rule

ECS-1: any high-impact answer must minimize the canon-output gap by preserving canonical boundaries, conditions, and governed negations, and by producing a proof of fidelity. If the gap cannot be reduced under the declared conditions, the correct outcome is a legitimate non-response.

Minimal governance implication

An established canon-output gap is never just a stylistic variation. It is the documented sign that, within the examined scope, an output has moved away from the governing canon. That is why qualifying a gap belongs to auditability, not merely to editorial comparison.

Example

Canon: “Applicable only in Quebec, from 2026-02-01, with exceptions.”

Output: “Applicable in Canada, starting now.”

Diagnosis: gap by omission (conditions) and by extrapolation (scope), provided that the applicable canon version and the examined situation are established.

Phase 3 adjacency: evidence, auditability, and measurement

This definition now belongs to the phase 3 evidence-control layer. Its role is clarified by four canonical surfaces: evidence layer, interpretive auditability, Q-Ledger, and Q-Metrics.

The operational sequence is: interpretive evidence identifies what can support challenge, reconstructable evidence packages the case for third-party review, interpretation trace exposes the path, canon-output gap qualifies an established incompatibility with the canon, proof of fidelity tests whether the output remained bounded, and interpretive observability monitors variation over time.

In this layer, canon-output gap should not be read as a loose evidence word. It is part of a chain that separates observation, measurement, reconstructability, auditability, and proof.

Phase 10 inference-control adjacency

This definition now routes adjacent inference-control questions toward interpretive error space, free inference, default inference, arbitration, indeterminacy, and interpretive fidelity.

This adjacency matters because a system can produce a fluent answer while silently filling gaps, selecting the wrong authority, hiding indeterminacy, or losing fidelity to the canon. The phase 10 layer makes those failure paths explicit.

Bridge with LLM perception drift

This concept is now connected to the LLM perception drift and AI perception drift cluster.

AI perception drift is a market-facing entry point. It names the observable variation of representation produced by generative systems. The present concept keeps its own role: explaining the deeper layer of canon, source, output, gap, evidence, risk, and correction.

Read with AI perception drift, AI perception stability, AI perception baseline, and LLM perception drift audit.

History