33 of 79 false positives on a code review benchmark were real defects
The Martian Code Review Bench offline set scored this run at 48.0% precision. A benchmark judge matches each comment against 173 human-written goldens, so a correct finding the annotator never wrote down is counted false by construction. All 79 false positives were opened in the source at the reviewed commit and given a verdict: 33 are real defects the gold set does not contain, 38 are genuinely wrong, 8 could not be settled from the code. Every verdict is published with its code and a pinned permalink, so any of them can be disputed with the same evidence in view.
Introducing Celmis
Self-hosted code intelligence over a symbol graph: what it does, what it refuses to claim, and how to run it on one machine.