Filter on lp_own >= 1.02*lp_empty (2% tolerance for negative reasoning lift)

Filter on lp_own >= 0.95*lp_empty (drop weak/negative reasoning lift)

Remove rows where reasoning hurt its own action (lp_own < lp_empty)

Strip <think> wrapper: answer_raw/messages are now plain THOUGHT: bodies

Use self-contained <think></think> prefix in answer_raw and messages

Add 14,965 repo-balanced rollouts with messages field (epoch-7 corpus)