cand_0004, v4's accepted G2-arm champion, reused as-is zero-shot, 5 trials per task, no further optimization, across all 136 v5 tasks. Every individual trial is drawn, so run-to-run stochasticity is visible rather than averaged away. Companion to heatmap_v5_g2_e1.html; narrative in results/v5/v5_g2_e1/summary.md.
Why this page exists. 56% of these tasks return a different reward across identical repeated trials (mean trial sd 0.051). A single task's before/after difference below roughly 0.09 is indistinguishable from noise at 5 trials, while the suite mean is resolved to ~0.02. A page showing only means would hide this. Every trial drawn here passed a per-trial audit of its on-disk record: 74 dead trials that cap-evolve had silently scored 0.0 were re-run, and the one remaining is excluded (platform-113, n=4).
open circle = one trial value (bigger = that value repeated) · filled = mean · line = min–max10-1010-12 which 68-task dropclick a column header to sort