$ python3 -I tools/experiments/charactereval_table_checks.py <eprint> <bundle>/results/table_checks.json
# started 2026-10-29T12:00:29+00:00  ended 2026-10-29T12:00:29+00:00  exit 0

source sha256 f98299737e75bfb7 (arXiv v2)
T1 Table 1 numbers [6811, 4564]; totals in text: ['11,376 examples (line 126)', '11,376 examples (line 168)']
T2 Overall printed {'Char-RM': 0.631, '1-shot': 0.362, '2-shot': 0.385, '3-shot': 0.375}; mean of 12 metrics {'Char-RM': 0.5611, '1-shot': 0.3676, '2-shot': 0.3852, '3-shot': 0.3951}; GPT-4 Overall increasing with shots: False
T4 90 dimension averages, max |printed - mean of printed metrics| 0.00067; 9 not equal to the rounded mean
T4 mark: cons Baichuan2-7B 3.827 bold=False underline=True (best 3.916, second 3.864)
T4 mark: cons Baichuan2-13B 3.864 bold=False underline=False (best 3.916, second 3.864)
MB 14 back-testing values, all k/216: True
NB baselines in text: ['10 baselines (line 530)']; Table 4 rows: 15