$ python3 -I tools/experiments/characterbench_audit.py <repo> <work>/table11.json <export> <work>/platform.json <bundle>/results/audit.json
# started 2026-10-27T14:03:13+00:00  ended 2026-10-27T14:03:49+00:00  exit 0

A data: 6494/6494 samples; 6494/6494 with every field equal; dropped 3; hash mismatches []
B prompts: role-play vs Table 11 0/6494 mismatches; judge input 0/6494
openai/gpt-4o-2024-05-13: judge input (real replies) 0/6460 mismatches; per-sample score diffs 0; all_eval stops on 0 outputs; max column diff 5.00e-03 (en.mr); cells differing at 2 decimals 0/28; unrounded {'zh': 8.881784197001252e-16, 'en': 8.881784197001252e-16}
deepseek-v4-1-flash-261017: judge input (real replies) 0/6488 mismatches; per-sample score diffs 0; all_eval stops on 0 outputs; max column diff 5.00e-03 (zh.fa); cells differing at 2 decimals 0/28; unrounded {'zh': 8.881784197001252e-16, 'en': 8.881784197001252e-16}
qwen-plus-character: judge input (real replies) 0/6474 mismatches; per-sample score diffs 0; all_eval stops on 0 outputs; max column diff 5.00e-03 (zh.bcpb); cells differing at 2 decimals 0/28; unrounded {'zh': 8.881784197001252e-16, 'en': 8.881784197001252e-16}
minimax/minimax-m2-her: judge input (real replies) 0/6494 mismatches; per-sample score diffs 0; all_eval stops on 0 outputs; max column diff 5.00e-03 (en.es); cells differing at 2 decimals 0/28; unrounded {'zh': 8.881784197001252e-16, 'en': 4.440892098500626e-16}
anthropic/claude-opus-5.5: judge input (real replies) 0/6476 mismatches; per-sample score diffs 0; all_eval stops on 0 outputs; max column diff 4.98e-03 (zh.er); cells differing at 2 decimals 0/28; unrounded {'zh': 8.881784197001252e-16, 'en': 8.881784197001252e-16}
openai/gpt-6-astra: judge input (real replies) 0/6494 mismatches; per-sample score diffs 0; all_eval stops on 0 outputs; max column diff 5.00e-03 (zh.bcpb); cells differing at 2 decimals 0/28; unrounded {'zh': 4.440892098500626e-16, 'en': 8.881784197001252e-16}
openai/gpt-4o-2024-05-13: judge input (real replies) 0/6462 mismatches; per-sample score diffs 0; all_eval stops on 0 outputs; max column diff 4.80e-03 (en.ms); cells differing at 2 decimals 0/28; unrounded {'zh': 4.440892098500626e-16, 'en': 8.881784197001252e-16}
doubao-seed-character-260628: judge input (real replies) 0/6494 mismatches; per-sample score diffs 0; all_eval stops on 0 outputs; max column diff 4.67e-03 (zh.bck); cells differing at 2 decimals 0/28; unrounded {'zh': 8.881784197001252e-16, 'en': 4.440892098500626e-16}
ranking zh: same=True
ranking en: same=True