The deployment that taught Yves to publish its own uncertainty.

A published satisfaction figure is a claim a regulator has to stand behind. A point estimate drawn from a corpus that may have been brigaded, presented with no stated uncertainty, gives a committee a very easy afternoon.

Wilson 95% confidence intervals went onto every share in the console, printed alongside the sample size in the same type size as the number, with thin cells marked as thin.

Every AI generation is archived verbatim with what was asked and what came back, so any figure in any board pack resolves to the exact computation that produced it.

The detection-rule catalogue was documented inside the product: twelve rules, each mapped to recognised techniques with the analyst response beside it. The console generates that page from the running system, so the documentation cannot drift from the code.

The board moved on from litigating the numbers and started arguing about what to do with them, which is the only outcome worth having.

Every rule set here became a standard for every deployment afterwards: quarantine and keep, archive every generation, state the interval, and hold a finding until the data supports it.