Most engines are trained to maximise responsiveness. Asked anything, they answer something. The hidden cost of that policy is that the rate of fabricated output rises smoothly with the difficulty of the question, and the user cannot tell when the threshold has been crossed. Quantm operates under the inverted rule: when the rule set is insufficient to derive an answer, the engine says so explicitly. It does not synthesise a confident-looking guess to fill the gap.
This discipline is what makes deterministic accuracy operationally meaningful. A refusal is a useful signal — it tells you exactly where the boundary of the engine's competence lies on this problem, so you can either reformulate the question, supply additional premises, or seek a different tool. A confident fabrication tells you nothing. It corrupts the downstream decision and hides the corruption behind fluent prose.
Refusal is also a security property. An engine that fabricates under pressure can be socially engineered into producing output it should not produce, simply by pressing harder. An engine that refuses cleanly removes that attack surface — there is no degraded mode in which the discipline weakens. The user gets either a verified answer or a clear refusal, and never anything in between. That contract is the deepest commitment the engine makes.