$ myla ~/how-we-measure
How we measure Myla
We score Myla against a fixed set of known cases every release: real repositories with planted issues and clean code mixed in. These are the numbers from the latest run, including what we already know we miss. No marketing rounding: percentages here round down, so a score never reads higher than it is.
Precision
100.0%
When Myla flags something, how often it is a real issue. The rest would be false alarms.
Recall
91.4%
Of the real issues in the set, how many Myla finds. The gap is what it misses.
The basis, so you can check it: recall is 53 of 58 planted issues found; precision is 53 of 53 flags correct. The denominators count findings, not cases.
The set has 93 cases, 2 of them known gaps that are excluded from both rates (below). Latest run: November 11, 2026 · commit 195c73d.
By category
What we miss
2 cases are a known gap: a check Myla does not pass yet. They stay in the set but are held OUT of the rates above (excluded, not scored as misses), so precision and recall describe the cases Myla is built to catch today, not these. Counted as misses, recall would be lower; they are listed here so the number above is not read as covering them. A finding that is not in the set is not measured at all.
Fix accuracy: of 2 one-tap fixes the harness ran, 2 produced the change we expected.