CORE GUIDE
Verification over introspection
When correctness matters, check the world or an independent artifact instead of asking the model whether its own reasoning was correct.
Mental model
The model's explanation of why it is right is another generated output. Verification uses independent evidence—a test, calculation, source, tool result or state check—that can contradict the model's claim.
Why it matters
Models can produce coherent rationales for wrong answers, and a second self-check can repeat the same hidden assumption. Introspection is useful for generating hypotheses, but it is weak as a correctness guarantee. Reliable systems therefore move important claims toward external evidence that has a different failure mode from the original generation.
01
Choose a verifier with independent evidence
Identify the claim that must be true, then ask what observable evidence would establish it. Code can be tested, calculations recomputed, citations checked against sources, database state queried and tool side effects reconciled. Use model critique when the property is qualitative, but calibrate it against human or deterministic checks rather than treating self-consistency as proof.
02
Example: did the migration preserve all rows?
A coding agent says the migration is safe and explains its reasoning. Verification instead runs the migration on a representative fixture, compares row counts and key invariants, and checks rollback behavior. The explanation may help review, but the executable evidence decides whether the claim is accepted.
Common failure modes
- Asking the same model 'are you sure?' and calling agreement verification.
- Using a fluent rationale as evidence that a factual claim is correct.
- Choosing a verifier that depends on the same untested assumption as the generator.
Engineering heuristics
- Translate important claims into observable postconditions.
- Prefer verifiers with failure modes independent from the generator.
- Use model critique as one signal, not the final authority, for consequential correctness.
Takeaways
- 01Self-explanation is not independent evidence.
- 02Verification should be able to prove the model wrong.
- 03The stronger the consequence, the stronger and more independent the verifier should be.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.