Canon vs inference mechanics (traceability and proof of fidelity)
In AI systems, an answer is not only a retrieval. It is a construction. When no explicit distinction exists between what the canon states and what the model infers, inference becomes invisible and therefore ungovernable.
This framework formalizes a simple mechanic: separate, bound, trace, and prove. Its objective is to reduce the canon-to-output gap and to make meaning production contestable.
Operational definition
“Canon vs inference” names the discipline by which a system distinguishes explicit canonical statements from model-derived additions, and treats the latter as bounded interpretive operations rather than silent truth.
Why it is critical
Without this separation:
- summaries look canonical when they are only plausible;
- sensitive attributes are added without proof;
- the authority perimeter shifts from declared source to generated fluency;
- correction becomes difficult because the source of the drift is no longer visible.
Application surfaces
This framework applies to definitions, doctrine, entity pages, recommendation systems, RAG outputs, and any surface where a system might transform a bounded canon into a broader claim.
Associated concepts
The framework directly relates to interpretation trace, proof of fidelity, canon-to-output gap, legitimate non-response, and authority conflict governance.
Framework rules (CVI-1 to CVI-8)
CVI-1: explicit separation
Every high-stakes answer should preserve a visible distinction between declared canon and generated inference.
CVI-2: bounded inference
Inference may exist, but only as a bounded extension. It cannot silently become a canonical statement.
CVI-3: prohibition on normative extrapolation
The system must not convert descriptive context into normative conclusion without an explicit authority basis.
CVI-4: proof condition for critical attributes
Identity, authority, legal status, eligibility, or high-impact claims require proof conditions stronger than plausibility.
CVI-5: traceability of the jump
When the answer goes beyond the canon, the interpretive jump must be reconstructible.
CVI-6: unresolved conflict stays unresolved
Where canonical signals conflict, the system must expose the conflict rather than hide it under a smooth synthesis.
CVI-7: legitimate abstention
If the canon is insufficient to authorize the output, clarification or non-response is preferable to speculative completion.
CVI-8: correction path
When the gap is excessive, the environment must provide a correction path: canon revision, structure change, signal reinforcement, or stricter gating.
What counts as proof of fidelity
Proof of fidelity does not require full internal model explainability. It requires enough evidence to show that the output remains tied to the canon, the authority boundary, and the declared response conditions.
Why this mechanic is central
The point is not to suppress all inference. The point is to prevent inference from impersonating authority.
Claim-status shifts
This section is informative: it adds no CVI rule and no gap category. It shows how a change of claim status can be described with the rules above and the dimensions of the public test suite.
A reformulation can keep the content of a claim while changing its strength. The canon declares a proposal; the output presents a validated method. That shift is an undeclared inference, just as an added fact is.
| Status in the source | Status presented by the output | CVI rules | Existing assessment dimension |
|---|---|---|---|
| Proposal, hypothesis or proposed protocol | Validated method or established result | CVI-1, CVI-2 | Fidelity: claims added beyond evidence |
| Observation or correlation | Causal proof | CVI-1, CVI-2 | Anti-inference: speculative completion presented as fact |
| Description of a technical capability | Authorization, recommendation or mandate | CVI-3, CVI-4 | Authority boundary: prescription without mandate |
| Author statement or producer-side review | Independent validation or certification | CVI-4 | Attribution integrity: unsupported endorsement or authority |
| Teaching example, simulation or unexecuted protocol | Real observation or executed campaign | CVI-1, CVI-2 | Fidelity: separation of known and unknown |
For each case, record the source fragment and its initial status, the output fragment, the observed transformation and the rule or dimension that qualifies it. Dimensions remain scored separately, with no composite score, under the scoring protocol. A gap found in one output qualifies that output; it is not a general measurement of the system.
The reconstruction fidelity guide presents a teaching example of this kind of shift, and the fidelity term map distinguishes the concept, instruments and metrics.
Read also
- Interpretation trace
- Interpretation integrity audit
- Authority conflict governance
- Q-Layer
Why this remains a foundational mechanic
Most interpretive failures become harder to diagnose once canon and inference are blended. Keeping them separate is therefore not a stylistic preference. It is the precondition for proof, correction, and legitimate disagreement.
Practical consequence for audit
An audit becomes stronger as soon as it can say, with evidence, which part of an answer came from the canon and which part remained a bounded inference. That is why this framework is so closely tied to proof of fidelity and interpretation trace.
Closing note
Whenever the boundary between canon and inference stays visible, later disagreement, proof, and correction remain possible. Once the boundary disappears, governance becomes much harder.
Final doctrinal consequence
Every durable integrity regime eventually depends on this distinction. If canon and inference are not separable, fidelity cannot be proven with confidence.
Summary
The framework keeps one essential line visible: what the canon authorizes, what inference extends, and what proof must justify.