adversarial-review

A cross-vendor code review uses a three-model panel of Claude Sonnet, Gemini, and GPT-5.4 to evaluate the same diff. It enforces diversity and tracks provenance across blind review, cross-examination, and adjudication phases to surface uncorrelated findings.

The workflow processes diffs through blind review, cross-examination, and adjudication phases. Findings are anonymised, pooled, and re-identified to ensure fair adjudication, ultimately producing a judge-ready packet for final decisions.

The adversarial code review process requires no external dependencies. It operates directly on diffs within a git repository, coordinating the three-model panel internally to enforce diversity, track provenance, and ensure reproducibility.

Cross-vendor code review improves uncorrelated issue detection by orchestrating three distinct model reviewers. By anonymising pooled findings and requiring explicit context, it prevents bias and surfaces diverse defects for the final judge packet.

Use a multi-model panel for git code review when you need to surface uncorrelated findings and require a judge-ready report. It is designed for reproducible workflows where tracking provenance, enforcing diversity, and safe adjudication are critical.