booting…
controller
learned (CEM-trained MLP)
scripted intercept
random weights
reset episode
run live A/B eval (5 seeds × 60 s)
Eval runs the SAME sim headlessly right now — nothing precomputed.
MLP 11→16→12→4 (tanh) trained headlessly in Node via Cross-Entropy Method · weights baked into weights.json · scene warm-started from proven conveyor build