Can deterministic static-analysis MCP tools improve audit results from non-deterministic AI? Three GPT-5.6 models reviewed the same Go codebase with and without them.
Glossary definition
Recall
Recall is the number of issues a method discovered out of the total number of issues. A recall of 30 of 32 means two issues were missed.
Recall is the number of issues a method discovered out of the total number of issues. A recall of 30 of 32 means the method discovered 30 issues and missed two.
Recall ignores noise: reporting incorrect findings costs nothing on this measure. That cost is measured by precision. Our private benchmark measures recall on unpublished audits, so no evaluated model has trained on the code or the reports.
Used in context
Articles using this term.
How good are frontier models at finding real vulnerabilities? Nine models measured against human auditors on five audits that never appeared in training data.