the ranking that pointed the wrong way

A sweep over the entire enumeration — every configuration the grid admitted, none skipped — returned nothing. Of the thousands of runs, not one survived the check against realized outcomes. On its own that is a routine negative, the kind of thing a programme reports and walks past on the way to the next question.

What made it worth writing down was why the plateau had looked promising for weeks. Candidates were ranked by the engine's own internal accounting. Compare that ranking with what actually happened and the ordering was worse than random: in the large majority of cases, the configurations the internal score liked best were the ones that came out worst once outcomes were marked at what they really were.

The score was not noisy. Noise produces a flat ranking — no signal, and no harm done. This ranking was anti-correlated with the thing it existed to predict, which means every hour spent optimizing against it moved the programme further from the answer, and the better the search worked, the further it went.

The response had two halves. The programme was declared verified-null instead of being continued into the next wave, and the realized check was promoted to a canonical, pinned gate that every later result has to pass with no drift between machines. What it proves is that instruments are tested rather than trusted, and that a metric which fails its own test is retired in public.