t5lab.arm_run

arm_run

Phase-2 planner-arm runner: for ARCH in {mvprop, highway, gppn, msp} run the three experiment pieces uniformly so the arms compare fairly: 1. RL-ONLY init probe — field value + move-score spread + init gradient norm at initialization (the mechanistic test for the vanishing-gradient pathology: is RL-only training even possible?). 2. DISTILLATION — supervised fit to the 4-connected wavefront teacher (masked MSE), the "does it learn to route at all" ceiling; saves planners/_distilled.eqx. 3. GENERALIZATION — optimal-move% + reach% across 8 unseen map types (the zero-shot metric).

PYTHONPATH=.:../../../FiedlerValueEstimation ARCH=highway GRID=32 STEPS=4000 $PY -m t5lab.arm_run