Module envelope 

What precision a detector can reach at our prevalence (D23).

Published detector numbers are measured on corpora that are roughly one-third positive, because they sample same-declaration parallel edits. Deployed against a mainline where bad merges are well under one percent, the same detector reports the same recall and a completely different precision — the false positives are drawn from a vastly larger pool of negatives.

That transformation is the difference between “60% recall at 43% precision, promising” and “one true alarm per 160”, so it should be computed rather than eyeballed. Recall and specificity are properties of the detector and carry across; precision is a property of the detector and the population and does not.

What the prevalence here actually is

The measured 0.004 is choir-queue::corpus’s revert-labelled bad-merge rate on rust-lang/rust — merges that history took back. That is not the same event as “merges containing a semantic conflict”: it misses defects fixed forward and includes reverts that were not defects. It is used because it is the only prevalence anyone has measured on a real mainline, and because the operational question is about flagging merges, not about conflicts in the abstract.

Structs

Detector

A detector’s prevalence-independent behaviour.

Functions

required_false_positive_rate

The false-positive rate a detector must not exceed to reach target_precision at prevalence with recall.