$ python3 -I tools/experiments/anchor_check.py --output /results/anchor.json # started 2026-10-28T16:43:33+00:00 ended 2026-10-28T16:43:53+00:00 exit 0 MiniMax public data: 34 of 55 pairs separated (Holm, alpha 0.05) gpt-5.1 > claude-sonnet-4.5 +2.02 Holm p 0.0231 minimax-m2-her > gemini-3-pro +2.58 Holm p 0.0149 claude-sonnet-4.5 > gemini-2.5-pro +3.56 Holm p 0.0000 gemini-2.5-pro > grok-4.1 +3.61 Holm p 0.0132 minimax-m2-her > claude-sonnet-4.5 +4.19 Holm p 0.0000 deepseek-v3.1 > deepseek-v3.2 +4.60 Holm p 0.0000 claude-sonnet-4.5 > deepseek-v3.1 +4.93 Holm p 0.0000 claude-opus-4.5-highthinking > gemini-2.5-pro +5.02 Holm p 0.0000 gemini-3-pro > gemini-2.5-pro +5.16 Holm p 0.0000 claude-opus-4.5-lowthinking > gemini-2.5-pro +5.25 Holm p 0.0000 gpt-5.1 > gemini-2.5-pro +5.58 Holm p 0.0000 gemini-2.5-pro > deepseek-v3.2 +5.98 Holm p 0.0000 claude-opus-4.5-highthinking > deepseek-v3.1 +6.40 Holm p 0.0000 doubao-1.5-pro > gemini-2.5-pro +6.47 Holm p 0.0000 gemini-3-pro > deepseek-v3.1 +6.54 Holm p 0.0000 claude-opus-4.5-lowthinking > deepseek-v3.1 +6.63 Holm p 0.0000 gpt-5.1 > deepseek-v3.1 +6.96 Holm p 0.0000 claude-sonnet-4.5 > grok-4.1 +7.17 Holm p 0.0000 minimax-m2-her > gemini-2.5-pro +7.74 Holm p 0.0000 doubao-1.5-pro > deepseek-v3.1 +7.84 Holm p 0.0000 claude-opus-4.5-highthinking > grok-4.1 +8.64 Holm p 0.0000 gemini-3-pro > grok-4.1 +8.78 Holm p 0.0000 claude-opus-4.5-lowthinking > grok-4.1 +8.87 Holm p 0.0000 minimax-m2-her > deepseek-v3.1 +9.12 Holm p 0.0000 gpt-5.1 > grok-4.1 +9.20 Holm p 0.0000 claude-sonnet-4.5 > deepseek-v3.2 +9.54 Holm p 0.0000 doubao-1.5-pro > grok-4.1 +10.08 Holm p 0.0000 claude-opus-4.5-highthinking > deepseek-v3.2 +11.00 Holm p 0.0000 gemini-3-pro > deepseek-v3.2 +11.14 Holm p 0.0000 claude-opus-4.5-lowthinking > deepseek-v3.2 +11.23 Holm p 0.0000 minimax-m2-her > grok-4.1 +11.36 Holm p 0.0000 gpt-5.1 > deepseek-v3.2 +11.56 Holm p 0.0000 doubao-1.5-pro > deepseek-v3.2 +12.45 Holm p 0.0000 minimax-m2-her > deepseek-v3.2 +13.72 Holm p 0.0000 == screening panel on MiniMax dialogues [judge]: undecided; 34 separated pairs covered, point direction kept 33/34, non-inferior (lower bound > -1.0) 32/34, gap retention 1.44 reversed, not significant claude-sonnet-4.5 > gemini-2.5-pro: MiniMax +3.35 (95 scenarios +3.56), candidate -0.65 [-6.01, +4.98], gap change -3.99 [-12.17, +3.25] (p 0.391, Holm 1.000), n=10 == official panel on MiniMax dialogues [judge]: undecided; 34 separated pairs covered, point direction kept 30/34, non-inferior (lower bound > -1.0) 27/34, gap retention 1.56 reversed, not significant claude-sonnet-4.5 > gemini-2.5-pro: MiniMax +6.39 (95 scenarios +3.56), candidate -5.23 [-14.84, +4.38], gap change -11.63 [-27.81, +4.56] (p 0.500, Holm 1.000), n=4 reversed, not significant doubao-1.5-pro > gemini-2.5-pro: MiniMax +13.34 (95 scenarios +6.47), candidate -0.16 [-8.71, +8.81], gap change -13.50 [-25.15, -1.84] (p 0.125, Holm 1.000), n=4 reversed, not significant gpt-5.1 > gemini-2.5-pro: MiniMax +10.53 (95 scenarios +5.58), candidate -1.31 [-8.07, +5.45], gap change -11.85 [-22.60, -1.09] (p 0.250, Holm 1.000), n=4 reversed, not significant minimax-m2-her > gemini-3-pro: MiniMax -5.04 (95 scenarios +2.58), candidate -2.84 [-5.93, +0.24], gap change +2.20 [-4.03, +8.43] (p 0.750, Holm 1.000), n=4 == platform v6 protocol (simulated user), screening panel [protocol]: undecided; 5 separated pairs covered, point direction kept 4/5, non-inferior (lower bound > -1.0) 2/5, gap retention 0.32 reversed, not significant gpt-5.1 > deepseek-v3.1: MiniMax +7.53 (95 scenarios +6.96), candidate -1.54 [-3.95, +1.49], gap change -9.06 [-13.87, -3.83] (p 0.014, Holm 0.068), n=10