experiment
Reducing noise in AI code review.
Testing whether parallel review, deduplication, independent judging, and confidence filtering can keep high-risk issues in front of a merge while reducing low-value feedback.
As AI-assisted changes grow, review can shift from missing risks to losing the real problems inside a large volume of feedback. This experiment focuses on producing fewer findings without reporting an incomplete review as a pass.
ReviewGate runs security, performance, logic, and AI-smell review in parallel. It then validates line locations, merges duplicate findings, applies an independent judge, and filters by confidence. Timeouts, context limits, and unfinished review return WARN.
Review high-risk defects first
Security, performance, logic, and AI-smell checks run by default. Pure style feedback stays out unless requested, while business rules and intent review are explicitly enabled.
Merge duplicates and judge each finding
Findings about the same location are merged before an independent judge checks the evidence. Low-confidence feedback is hidden by default but remains available for inspection.
The core path works with real pull requests and CI
ReviewGate provides CLI and CI integrations. An incomplete review returns WARN instead of treating missing evidence as PASS. Public evaluation records continue to be used to calibrate false positives and missed issues.