A generative system should not be evaluated only on whether a statement is “true enough”. The real question is how far the output has moved away from the canon, on which dimensions, and with what operational consequences.
Operational definition
The canon-output gap is the measurable distance between a canonical source and the output reconstructed by an AI system. That distance can be lexical, perimeter-based, normative, authoritative, or intentional. The purpose of the map is to qualify distortion before it hardens into a stable public interpretation.
Why the debate must shift from “truth” to distortion
In an interpreted web, outputs are compressed, reformulated, and normalized. A response can look plausible while still changing the perimeter, hierarchy, tone, or authority of the original statement. The strategic problem is therefore not only factual error, but the loss of fidelity between canon and synthesis.
Dimensions of the gap
- Lexical gap: key terms are replaced by near-synonyms that alter the intended meaning.
- Perimeter gap: the output extends or narrows the scope of what is actually covered.
- Normative gap: what was conditional, optional, or uncertain is reformulated as stable or required.
- Authority gap: the output upgrades, downgrades, or replaces the actual source hierarchy.
- Intentional gap: the output changes the practical implication or expected use of the source.
Practical measurement protocol
- Start from a stable canonical page or definition, not from a cluster of secondary mentions.
- Test the same object across multiple prompts and models in order to observe recurrent distortion.
- Classify each deviation by dimension rather than collapsing everything into “hallucination”.
- Measure severity according to consequences: identity fusion, perimeter drift, false obligation, or authority conflict.
- Escalate only the gaps that change governability, not every stylistic variation.
What this map prevents
- Treating plausible reformulation as harmless when it actually changes the decision surface.
- Debating “truth” without identifying which layer of the canon has drifted.
- Correcting wording while leaving the structural source of distortion intact.
- Allowing repeated approximation to become a fixed public attribute.
