Where this leaves you

The model that writes the answer now reads three laws chosen by a model that read thirty-two, and in your pipeline, when that model fails, it reads the eight it read before. Both model jobs run on Haiku, with a five-second leash.

Three laws are the right three more often than eight were. They are still not always enough, and the reason is in the shape of a code: the duty is in one subsection, the fine in the next, and the way out in the one after. Retrieval found one of them. The rest is a lookup.