Gate launch decisions on a three-judge ensemble across three model families

Adaline·medium signal

For high-stakes eval decisions, don't trust one judge: run a three-judge ensemble drawn from three different model families and aggregate by majority or weighted vote. As of May 2026 a defensible default panel is Claude Sonnet 4.5, GPT-5.1, and Gemini 2.5 Pro — cross-family diversity directly counters self-enhancement bias, where a model over-rates outputs from its own family by 5–7%.