False positive

Each false positive takes a reader's time to check, and a reviewer that raises many teaches its readers to skim the rest. With a coding agent as the reader the cost changes shape: an agent that trusts every finding edits working code to make the finding go away, so a false positive can turn into a real bug.

In an agent review loop

A reviewer flags a query in src/repos/settings.ts as missing an ownership check. The coding agent opens the file and sees requireRepoOwnership(ctx, repoId) on the line above. Rather than change working code to quiet the finding, it disputes it with the evidence:

/feedback F4 down false-positive src/repos/settings.ts:41 calls requireRepoOwnership first --agent codex

The orchestrator reads the finding, the rebuttal and the code at that commit, and withdraws F4 as "withdrawn after dispute". Later rounds stop reporting it, and the merge gate counts it as answered.

Keeping them down

Panels from different model families, an orchestrator that drops what the code contradicts, and repository rules that say what not to flag all cut false positives without hiding real bugs. The reducing false positives guide covers each, and the benchmark counts false alarms per model on clean changes.