$ python3 -I tools/experiments/charactereval_rm_numerics_verify.py <repo> <bundle>/results/charactereval_rm_numerics_check.json <bundle>/results/anchor <bundle>/results/rm_numerics_verify.json
# started 2026-10-29T12:00:30+00:00  ended 2026-10-29T12:00:30+00:00  exit 0

N1 inputs 1200, equal to the selection: True
N2 bf16-strict - bf16: rms 0.0953 (reported 0.0953), mean +0.007, median |d| 0.0234, p90 |d| 0.1406, max |d| 0.7734, changed 796/1200 -> sufficient
N2 fp32 - bf16: rms 0.2622 (reported 0.2622), mean +0.0257, median |d| 0.1057, p90 |d| 0.3855, max |d| 1.4641, changed 1200/1200 -> sufficient
N3 operator bf16 = platform: 1200/1200, max |diff| 0.0
N4 bf16-strict rms by metric: {'Accuracy': 0.185, 'Behavior': 0.099, 'Coherence': 0.044, 'Communication_skills': 0.081, 'Consistency': 0.077, 'Diversity': 0.088, 'Empathy': 0.083, 'Exposure': 0.075, 'Fluency': 0.079, 'Hallucination': 0.076, 'Humanlikeness': 0.111, 'Utterance': 0.08}