Parsec v4 heatmap — v4_g2_e1
All 34 v4 Parsec scenarios · cap-evolve zero-shot baseline (n=5) vs v4_g2_e1: a single global-scope bundle evolved across the whole task set (seed + 6 candidates, n=5 per eval, all 34 tasks every iteration). Unlike T2's per-task bundles, G2 has no per-task "optimized/not optimized" split — every candidate is evaluated on every task. Optimizer selected cand_0004 as best_id; per-task "best" below picks each row's own max, which can differ from the run's overall best_id.
Regression and challenge tranches are never pooled into one mean — the challenge tasks are designed to score below 1.0, so mixing them into the regression mean would misread deliberate difficulty as regression. Every mean below shows its own n=; different columns use different trial counts and are not directly comparable without it.
0.0
0.5
1.0
regression expected 1.000
challenge expected <1.000 by design
best < 1.0
Click a column header to sort.