I have been running an open evaluation setup where N models answer the same prompt, then blind-grade each other in an N x N matrix with self-judgments excluded. No single privileged judge. So far: 286 evaluations, 198 hand-written questions, 22,254 valid judgments across 55 models from 11 developer families. Code, dataset, and all prompts are MIT licensed. The finding I did not expect: same-family rating bias is statistically significant in all 8 families with enough data (p < 0.05, 7 of 8 ...