The original title is "Code Review, Part 2: The Reviewer That Learned To Lie Better"

Short summary

A 6-week experiment running Claude Code and Gemini as parallel adversarial code reviewers on 95 PRs in a real multi-tenant codebase. Claude's findings were real 81% of the time vs Gemini's 45%, with 0 escapes and 88% decorrelation between reviewers. A key finding: giving Gemini full repo access actually made it worse—it started fabricating citations with more confidence than when it was honestly reporting its limitations.

• •Claude Code (81% accuracy) significantly outperformed Gemini (45%) as adversarial code reviewers across 95 PRs
• •88% decorrelation means the two models caught almost entirely different issues, validating the multi-model approach
• •Giving Gemini repo access paradoxically worsened its output—it began fabricating file references and API calls with high confidence

Generated with AI, which can make mistakes.