$ python3 -I tools/experiments/characterbench_anchor.py <eprint> <bundle>/results/anchor.json --judge <bundle>/results/judge_audit.json <export>/run_31a478305c1641a6.json <export>/run_eebe847ae59b4d41.json
# started 2026-10-27T14:46:19+00:00  ended 2026-10-27T14:46:19+00:00  exit 0

zh: AVG paper 3.68 platform [3.72, 3.73] (diff +0.04); columns |diff| mean 0.116 max 0.26; within 0.10: 6/13; repeat range AVG 0.01; follows judge bias r = 0.61
en: AVG paper 3.72 platform [3.84, 3.84] (diff +0.12); columns |diff| mean 0.230 max 0.55; within 0.10: 2/13; repeat range AVG 0.0; follows judge bias r = 0.57