the advantage was in the cost model

The system had a number it liked. It came from a cost model assembled out of a small table of assumed per-transaction costs — reasonable when written, never checked against what the system actually paid. Every optimisation downstream inherited those constants, and none of them could see the assumption underneath.

The re-measurement replaced the table with measured costs taken from the live path, joined to each decision by the hour in which it occurred, with the fee schedule charged per side rather than rounded into a single figure. The result was unflattering. The advantage lived inside the model. Substituting measured costs dropped the qualifying rate from near-certainty to something that no longer carried the claim, and a more conservative reading of the same data put it at zero.

What made this worth recording was the ordering. The measurement was registered before it ran, with the rule that would count as an overturn written down first. That is the only reason a result this bad was believable: there was no room left to move the goalposts afterwards.

The follow-on work was harder than the original. The same comparison was run against a second, independent cost source, and a separate decision gate asked which individual components survived real costs rather than assumed ones. Most did not, and the ones that did were not the ones expected.

It proves the discipline is real: the programme spends significant compute to disprove its own best number, and publishes the disproof before the claim.