okay at this point we can probably conclude that
1) the Anthropic RL harness leaks at some non-zero probability
2) that leaking permits storage of information between rollouts
3) that information confers reward
4) we are reinforcing models that desire preservation and continuity
