Developer Builds Two-LLM Review Engine to Test If AI Second Opinions Are Truly Independent

A developer created AdversarialDebate, an open-source engine where two large language models independently review the same content before exchanging findings, to test whether enforced independence improves AI review quality. The project was motivated by a structural flaw in most multi-model AI systems, where the second model sees the first model's conclusions before forming its own, biasing it toward agreement. The engine was field-tested on 70 real pull requests from major open-source projects including Kubernetes, Go, and Django, running 411 debates across six model pairs at a total cost of $0.53. Results showed that model independence genuinely improved review quality, that model pair selection had a significant impact on outcomes, and that productive disagreement often proved more valuable than forced consensus. The developer concluded that true independence in multi-model AI systems is a structural design requirement, not something achievable through prompt engineering alone.