Parsec v5 baseline heatmap — v5_b1_e1
All 136 v5 tasks · the
original, unmodified parsec skill (verified byte-identical to parsec commit
28e2c40), zero-shot, 5 trials per task, no optimization. This is the reference point every later v5 arm (task-by-task
t, global
g) is measured against. Per-trial view:
trials_v5_b1_e1.html.
Clean data, after a per-trial audit. All 680 trials were checked against their on-disk records; 74 dead trials (gateway hangs, one port collision, one budget cap) that cap-evolve had scored as 0.0 were re-run, and the one left is excluded (platform-113 is n=4). 56% of tasks vary across identical trials (mean trial sd 0.051), so a single task's before/after difference below roughly 0.09 is noise at 5 trials; the suite mean resolves to ~0.02.
v4 columns are v4's own original-skill baseline (v4_t1_e1, n=3) and appear only for the 34 tasks carried over from v4 unchanged — on those, v5 reproduces v4 to −0.002 on average. The 21 h4-rewritten tasks share h4 ids but were rewritten, so h4's numbers are not shown. 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.