a perfect score at the ceiling proves nothing

A campaign sweeps a large grid of configurations and ranks them on a computed score. Near the top of the grid, the score stops being informative, because the quantity being scored is bounded — a proportion cannot exceed one — and any design that would almost always succeed saturates it.

One selected configuration came back at the maximum. The pipeline refused to read that as a verdict. A guard fired and labelled the result degenerate at the ceiling: the score is pinned at its upper bound, so it cannot distinguish a design that is genuinely excellent from one that merely crossed the bound, and reporting it as a clean result would be reporting the shape of the metric rather than the quality of the design. The record states the guard's decision next to the score, and the recommendation was carried by other evidence instead.

The same pass turned up an accounting defect: a family of rows carried charges for work that had never been executed, because a configuration excluded one engine and the charge table had not been told. That was corrected, and the correction sits in the record rather than being quietly patched away.

The cost is that a headline number lost its headline. The programme gave up a clean story about a perfect score and kept the smaller, defensible one.

What it proves: the measuring apparatus is subject to the same review as the thing measured. A number that cannot fail is not evidence, and a programme that knows the difference is a programme whose other numbers mean something.