Recipe ablation · 32×32 / 10 agents · 4 recipes × 3 maps × 3 seeds (36 runs)

Recipe ablation — which ingredient actually earns its place?

What it tests: a controlled ablation that toggles three ingredients of the best recipe one at a time, on three maps (open · rooms · crowded), three seeds each. The four arms are role_soft_bump (explorer/relay roles + soft-λ connectivity + a bump-explore reward term), base_soft_bump (drop the roles → homogeneous), role_lag_bump (swap the soft penalty for a learned-Lagrangian one), and role_soft_flat (drop the bump term → flat reward). Reading the four side by side isolates what each piece is worth.

Key finding Measured live from the rendered tiles below. role + bump wins, and Lagrangian vs soft is essentially a tie. The single biggest lever is the bump-explore term: removing it (role_soft_flat) is the floor. The explorer/relay role split helps most on open and crowded. Three seeds per cell.
Per recipe — mean of all maps × seeds · rollout coverage = final % of free cells visited (100-step CPU re-roll)
RecipeWhat it changesMean coveragevs floorruns
computing from the rendered rollouts…

The run list is grouped by recipe, then map — expand a recipe to compare it across terrain. The table is computed live from the rendered rollouts, so it self-corrects on every re-render.

t = 0