Canon-output gap as a measure of LLM perception drift
LLM perception drift becomes measurable when a dated output can be compared with a dated reference state. The canon-output gap asks which properties, boundaries and authority roles were preserved or transformed, not whether the answer merely “looks good.”
Two bounded objects
The canon contains claims the organization or another competent source may legitimately establish. The output contains generated text, context, citations and uncertainty. Neither should be idealized.
Canon may be incomplete or promotional. Output may include legitimate external evidence. The gap is not simply “different from the official site.”
Comparison dimensions
| Dimension | Question | Example gap |
|---|---|---|
| Identity | Is the correct entity named? | Namesake confusion |
| Category | Is the primary role preserved? | Specialist reduced to generalist |
| Scope | Are offer and exclusions respected? | Out-of-market service added |
| Time | Is current state distinguished? | Old product shown as active |
| Relations | Are affiliations and competitors correct? | Subsidiary turned into partner |
| Authority | Does a competent source govern the claim? | Old directory against current policy |
| Non-implication | Does output exceed evidence? | Expertise turned into leadership |
| Reputation | Are evaluations attributed? | Opinion presented as consensus |
Coding method
Classify each invariant as preserved, partial, displaced, contradictory, legitimate external or not observable. This prevents treating every omission as error and every criticism as drift.
Example
The canon describes a firm as a B2B specialist in governing AI answers. An output calls it an SEO agency using AI tools. Name and some skills are preserved, but category, hierarchy and audience are displaced. Sentiment is neutral and no spectacular fact is invented.
The useful measure is not sentiment. It is invariant preservation, gap severity and repetition across intents.
Aggregates and critical incidents
A global index may help tracking, but must not hide rare severe cases. Report dimension aggregates, prompt-family distribution, critical incidents, cross-model and cross-language divergence, structuring sources and confidence.
Temporal comparison
Drift requires comparable windows. Measure what appears, disappears, repeats or changes in severity. Model, browsing or prompt changes may invalidate comparison; protocol versions must remain visible.
Limits
The canon-output gap does not prove market acceptance, directly measure public image or brand equity, or establish the cause of output. It provides an evidence object: what was established, what was produced and how they diverged.