TrueMend's wrong-match rate on real Python code: 4.5%, beating Ruff's 6.1% — independently blind-adjudicated, not self-reported.
Ruff is the fastest, most widely adopted Python linter available. If TrueMend can't hold its own against Ruff on real code, no amount of AI-focused framing matters. So we measured it directly.
The claim
A finding is only as good as its precision. A tool that reports 500 issues but gets half of them wrong isn't saving anyone time — it's adding a triage problem. We ran TrueMend and Ruff against the same real, unmodified Python codebase and asked one question per finding: is this claim factually correct?
Method — blind two-axis adjudication
- An independent judge model rules each finding correct or wrong. It doesn't know which tool produced the finding, and it isn't told what the "right" answer should be.
- Same codebase, same commit, for both tools.
- A separate Goodhart audit specifically checked that no TrueMend detector had been hollowed out, disabled, or deleted just to make this number look better — precision has to come from getting findings right, not from firing less.
Result
TrueMend: 4.5% wrong-match. Ruff: 6.1% wrong-match. TrueMend wins, independently confirmed — not a preliminary or self-reported number. Getting here took two real remediation waves fixing 6 concrete precision bugs, not a single lucky measurement.
Why this one is different from our other two benchmarks
We're publishing this number as confirmed, not preliminary, because it went through full independent blind adjudication — the same rigor a security audit would demand. Our TypeScript/React and Rust benchmarks (below) are real wins too, but self-measured this past week and explicitly labeled preliminary until they get the same independent adjudication treatment. We think showing the difference in confidence level is more useful to you than pretending every number carries equal weight.