Silent error is the most dangerous failure mode in any reasoning system. The output looks right. The surface form is fluent. The numbers are well-formed. And the derivation underneath is unsound. A reviewer reading only the output cannot detect the failure, because there is nothing in the surface to detect. The verification pass is the architectural defence against that class of failure.
A Quantm verification pass takes the candidate output and re-derives it along a structurally independent path. 'Structurally independent' is the load-bearing phrase. Re-running the same derivation catches transient noise but not systematic error. Re-deriving along a different decomposition, a different inference primitive, a different limiting case, or a different unit-analysis bound exposes the kind of silent error that can survive a single derivation. When two independent paths agree to within tolerance, the joint probability of a coherent silent error is low enough to commit. When they disagree, the engine refuses rather than picking one.
The pass is not free. It costs synthesis budget — which is why higher tiers buy more verification depth, not more output length. But the cost is structural, not optional. A single-path answer is honest about its verification depth; a multi-path answer is honest about its hardness. Either way, the user knows what they are holding, which is the entire point.