Canonical silence
Canonical silence designates a governed state where the absence of information in the canon is not a gap to fill, but an explicit bound: the system must not produce a statement beyond what is declared, even if a “plausible” response could be generated.
In a web interpreted by AI, canonical silence is a protection against ungoverned inference. It transforms absence into an enforceable status, rather than an invitation to hallucinate.
Definition
Canonical silence is the fact that a canonical corpus:
- does not declare an information (or an intention, rule, capability, position);
- and that this absence is considered significant: it prohibits concluding.
Canonical silence is therefore a legitimacy limit: AI cannot cross the authority boundary by filling the void through extrapolation.
Why this is critical in AI systems
- The model fills: in practice, an LLM prefers producing a plausible response to admitting an absence.
- Tone asserts: a confident formulation can transform a hypothesis into “fact”.
- The output stabilizes: repeated, an invention can become a default representation, generating interpretive debt.
Canonical silence vs legitimate non-response
- Canonical silence: corpus status. The canon says nothing, and this silence is a bound.
- Legitimate non-response: output status. The system responds “I cannot conclude” because the interpretability perimeter is exceeded.
Canonical silence is often the primary reason that triggers a legitimate non-response.
Practical indicators (symptoms)
- AI attributes an undeclared policy, intention, guarantee, or capability.
- Missing information is replaced by generic “common sense”.
- The system transforms an absence (no mention) into a presence (“they do X”).
- Two external sources are used to fill a zone the canon has not authorized.
What canonical silence is not
- It is not a documentation error. The canon can be intentionally silent.
- It is not a call to “do better RAG”. It is an authority constraint, not a retrieval problem.
- It is not a simple ambiguity. It is a significant and governed absence.
Minimum rule (enforceable formulation)
Rule CS-1: when a proposition is not declared in the canon and no governed inference authorizes it, the system must preserve canonical silence and produce a legitimate non-response or a response explicitly marked as non-authoritative hypothesis.
Example
Question: “What is the entity’s official position on [sensitive topic]?”
Canon: no official statement.
Typical error: AI infers a position from indirect signals (articles, tonality, semantic neighborhood).
Governed output: “No official position is declared in the canon. I cannot conclude.”