[m1rl0k]

Replaces the October 5 Sonnet 4.5 figures with the October 20 Claude Sonnet 5.5 paired A/B (80 runs, 40 tasks x 2 seeds, same two repositories):

• Tool calls per task: 14.2 -> 10.4 (-26%, p < 0.001)
• Agent turns per task: 15.2 -> 11.4 (-25%, p < 0.001)
• Model cost per task: $0.118 -> $0.094 (-20%, p = 0.02)
• Accuracy held level (60.5% vs 61.3%); tokens -18% not significant, so not claimed

Task-clustered bootstrap CIs, task-level sign-flip permutation p-values. The earlier Sonnet 4.5 accuracy result is kept in the details text.

🤖 Generated with Claude Code