When you have an auto-catalyzing system, don't write down the answer, but the process that made it.

That way, if the context changes and the answer should change, the process can rediscover it automatically.

This was true with distilled ranking signals in Search.

But it's also true in the world of LLMs.