Step-100 checkpoint from arc_pure_rl_icl_mb64_200s run (200-step plan, killed at step ~104 due to policy collapse; step-60/80 eval was top1_em=0.2812 vs base 0.1875; step-100 eval regressed to 0.1875). 15GB full base model. e255680 verified HerrHruby commited on 16 days ago