Inspecting compiled semantics
No matching nodes
Choose up to two semantic kinds. One reveals its real traffic; two compare only direct authored relationships.
Choose a kind to inspect its nodes and touching relationships.
Semantic radar needs more MAP space.
How do we check polywatch against bugs a project actually had?
A reviewer that sounds convincing can still be useless. The first test is a project's own history, where later fixes show which commits were wrong.
The git replay reads the project's history. For every commit whose subject starts with fix, it blames the lines the fix changed and marks the commits that wrote them as buggy (the SZZ method). Commits that no fix ever touched, and that are old enough, are marked clean.
Each sampled commit becomes a job for the same worker the plugin uses, so the replay measures the shipped pipeline.
After the run, Opus compares each buggy commit's findings with the fix diff and says whether any finding names the bug that was fixed. The answer key is noisy: an unfixed bug looks clean. On BaanBaan this replay showed that the reviewer's verdict carries no information, and that its specific claims sometimes do.