Datasets:
Task-level bootstrap of the published results: the top two are statistically tied
Thanks for publishing all 600 runs with their grades; it makes the results checkable. I resampled the 50 tasks with replacement (20,000 draws, seed 2026), using the per-task trials_passed in reports/meridian/results-2026-09-27.json, and recomputed each model's mean pass rate. deepseek-v4p1-flash comes out first in 56% of resamples and gpt-6-sol in 43%; the difference between them is +1.3 points with a 95% interval of about -10 to +13. Different summaries also order the two differently: deepseek leads on mean pass rate (60.0% vs 58.7%) and pass@3 (38 vs 35), while sol leads on pass^3 (24 vs 23) and wins more head-to-head tasks (12 vs 9, with 29 ties). By contrast, deepseek against glm-5p3-flash is a real gap (better on 17 tasks, worse on 4, sign test p of about 0.007). Showing intervals, or marking statistically tied models, next to the ranking would stop readers from reading the top two as an order.