the summary table that reported a loss as a gain

An independent reviewer went through a family of published result tables looking for internal inconsistency rather than wrong arithmetic. They found one that mattered. The worst-case column carried no sign. A minimum had been formatted through a path that treated the value as a magnitude, so the worst interval in a sample — an actual loss — printed exactly as though it were a gain, and a sample that was entirely positive printed identically. The two cases were indistinguishable in the table.

Two invariants caught it, both checkable from the tables alone: the worst interval must not exceed the low-percentile row, and it must not exceed the worst single run's decline. The column violated the first in most rows and the second in nearly all of them.

The fix was narrow, and its narrowness was the point. The distribution was untouched; not one trial changed. The generator was corrected to emit the signed minimum, the affected deliverables were re-printed, and the diff was verified line by line to contain only headers and that one column. A regression test now feeds the emitter both an entirely lossy walk and an entirely profitable one and checks the sign survives both.

The honest framing is that the numbers were right and the label was wrong — the failure mode that survives longest, because every individual figure looks plausible.

It proves the review aims at how results are described, not only at whether they compute.