the prudence filter kept the wrong half

A rule was proposed that sounds like discipline: keep an execution only if its measured cost is within a small multiple of the cost the model assumed for it, and drop the rest. The premise is that costlier-than-modelled work is work nobody understood, so removing it leaves a cleaner book.

Measured, the rule removed between a quarter and a third of the record. It also removed, with almost perfect selectivity, the members that had been carrying the result: the strongest member lost nearly all of its entries and another lost every one, while the members already known to be the weakest passed through untouched. The property the gate selected for was real and measurable. It was simply not the property that mattered.

The comparison was then run both ways under the same conventions, gated and ungated. Neither cleared the standing admission test, and what the gate changed was the composition of the survivors, leaving a set chosen for cheapness rather than a set chosen for anything the programme wanted.

The rule was not adopted, and the reason is in the record rather than implied: a filter that is a proxy for quality has to be checked against quality, not against its own premise. The measurement was registered before it ran, one charge was spent on it, and what came back is that a plausible control can be anti-correlated with the objective while looking like caution.

What it proves is that the programme measures its own safeguards before trusting them.