two equivalent conditions, one of which fires on nothing

A condition can be written two ways that are equivalent for every number and behave differently for the value that is not a number. The reference implementation asked whether the magnitude of a statistic had reached a threshold, written positively — so an undefined statistic compared false and nothing fired. The fast engine asked the negation, whether the statistic had not failed to reach the threshold, and an undefined value passed straight through and fired.

In an environment with no missing inputs this never occurs. Here it occurred on windows where the underlying statistic was flat: a division of zero by zero, produced legitimately by a quiet stretch, which the negation read as a signal.

What made it findable was scale. The disagreement appeared as one missing record and two extra among more than a million, only on the full-length run; every smaller test was clean, and every smaller test therefore certified a defect.

The fix is one comparison inverted. The durable output is a rule applied to every similar predicate: never optimise a guard into its negation, because the negation admits exactly the cases the guard existed to exclude.

There is a second lesson in the shape of the evidence. The extra records were the interesting half. A missing record looks like a gap and can be argued about; an extra record is proof that something executed which should not have.

It proves the engine is held to the reference record by record, not on aggregate agreement.