Make the Model Commit

A language model will produce a fluent, confident explanation of a scene that cannot physically exist — and an equally fluent one of a scene that can. The output surface is too cheap. Prose never has to commit to a constraint that something else can check, so a wrong answer costs the model nothing but a slightly different paragraph.

In one project I stopped asking the model to explain and started asking it to commit: emit a structured program — declared contacts, supports, centre-of-mass claims — and hand it to an SMT solver. Now a wrong claim is not a softly-worded sentence; it is an UNSAT. Either the centre of mass is over the support polygon or it is not. The model proposes; the checker disposes.

Anywhere the answer has checkable structure — geometry, arithmetic, schedules, types, proofs, API contracts — you can push the model to emit that structure and verify it externally. This is the spine of the neuro-symbolic revival, and it is the most reliable way I know to make machine reasoning gradeable instead of merely persuasive. The verifier turns confidence into a testable claim.

This only works where you can formalise the check, and most interesting questions resist formalisation. The space of problems with a clean external verifier is smaller than we would like. But for the slice that has one, “make the model commit to a verifiable artifact” beats “trust the prose” every single time — and it is worth designing problems so that more of them fall into that slice.