v1 fine-tune vs v2 FlowDAgger steering — offline gate, 100 held-out states
π0.5 LoRA · frozen ckpt 29999 · UR10e drug-sort · chunk-recon MSE vs IK expert · stock ≙ v1 · steered ≙ v2 · offline proxy — Stage B pending
APer-state · below the diagonal, steering wins
BSame frozen model · 4.3× lower error
CSteering fades along the task
DLosses are ties · wins are large