the same answer on a different machine

A large sweep produced a map of candidate configurations, and the map was produced once, on one machine, by one program. That is exactly the situation in which a subtle dependency on the machine — an ordering, a threading choice, a difference in how the arithmetic rounds — hides inside results that look fine.

The verification was deliberately not a re-read of the existing output. Two substantial slices of the map were recomputed from the raw inputs, on a second machine, from scratch, with no reuse of the first machine's intermediate products. The two outputs were then compared byte for byte.

They matched. Every record of both slices was identical to the first machine's. That is a narrow claim, and a load-bearing one: it rules out the class of defect where the answer depends on where it was computed, and it rules out a family of silent nondeterminism that would otherwise surface much later, at the worst possible moment.

The cost is blunt. This is the most expensive kind of check available: it doubles the compute for whatever it covers, and it can only ever cover a sample, so it buys confidence in a mechanism rather than a guarantee about every row. The slices were chosen to be large enough that a machine-dependent defect would have had to hide in a very peculiar way to survive them.

What it proves: the results are a property of the method and the inputs, not of the machine that ran them.