Parsec v5 G2-arm heatmap — v5_g2_e1
All 136 v5 tasks ·
cand_0004, v4's accepted G2-arm champion (
artifacts/v4/v4_g_e1/best, verified byte-identical to the GitHub-hosted copy), reused as-is, zero-shot, 5 trials per task, no further optimization. Compared against
v5_b1_e1 (the original skill, same tasks, same settings). Per-trial view:
trials_v5_g2_e1.html.
Clean data, after a per-trial audit. All 680 trials were individually checked against their on-disk records. 4 tasks never reached 5 clean trials in any single run across repeated resubmission (recurring gateway hangs); for those, every error-free trial across every attempt on disk was pooled and 5 were drawn at random for the mean (see results/v5/v5_g2_e1/summary.md). 0 of 680 final trials carry an unresolved infra problem.
v5_b1 columns are v5_b1_e1's own reward on the same task (the original, unmodified skill) — the real comparison for this arm. No group clears the ~0.09 per-task noise floor measured in v5_b1_e1; MEAN rows pool the tasks named in them.
0.0
0.5
1.0
10-1010-12 which 68-task drop · v4-34 carried over from v4 unchanged · h4-rewritten h4 id, rewritten task
Click a column header to sort.