A probabilistic synthesiser optimises for the likelihood that the next token sounds correct given everything that came before. That objective produces output which is locally smooth and globally indefensible: no individual claim has a defensible derivation, and the chain of claims cannot be audited because there is no chain — there is only a sequence of plausible continuations. For a journal entry or a creative draft, this is fine. For an engineering result, it is the wrong tool.
Engineering operates under a stricter contract. A specification is satisfied or it is not. A load is within bound or it is not. A reconciliation balances or it does not. There is no continuum of plausibility on which 'almost correct' earns partial credit. A probabilistic system cannot tell you why its answer is what it is, because the system was never designed to derive — it was designed to predict. Asking it for a derivation produces a synthetic narrative that resembles a derivation without being one.
Quantm replaces prediction with derivation as the underlying operation. Every output is the unique consequence of an explicit rule chain applied to explicit premises. The chain is recorded, the premises are typed, and the rules are sound. If the answer is wrong, the wrongness is locatable. If the answer is right, the rightness is auditable. That structural choice is the only thing that makes the engine usable on problems where being almost right is the same as being entirely wrong.