Bjørn Johansen6 min readintermediate
Three AI models review every merge request. Then they argue.
Summary
A GitLab CI gate runs three isolated LLMs (Claude, GPT, Grok) on every merge request, forces them to exchange findings with code‑backed evidence, and decides mergeability via a deterministic voting function that only blocks on findings with independent confirmation.
- Run each LLM reviewer in complete isolation (no shared context) to get independent opinions.
- After the first pass, feed anonymized findings to the other models; they must confirm or refute with concrete code evidence.
- Only findings that receive independent confirmation (or survive a capped third‑round debate) can block a merge; no single model can unilaterally approve or reject.
- The final decision is a pure function over a stance matrix (confirm/refute, severity, fixability), ensuring deterministic outcomes.
As AI‑generated code proliferates, human reviewers become a bottleneck. This multi‑model, evidence‑driven gate provides a scalable, reproducible safety net without relying on any single LLM’s whims.
6/10
