Conclusion

The split between what will be superseded and what will not runs along a clean line.

The two model identifiers in setup.ts will be superseded, probably within a year, and the Maryland dates in corpus.ts sooner than that. Both are the kind of fact this pipeline was built to carry rather than to hide: one edit in one file, and a date under every answer.

The parameters will move. output_config, effort, the shape of a structured reply, the name of the embedding model and the size of its vectors: each will be renamed, become a default, or be replaced by something better, and the listings will need a line changed where they were.

The limits will move too. Five seconds for a side job, 600 characters for an answer, 24 or 32 laws in a pool, 900 characters in a chunk, 0.5 for a neighbour to count. Every one of those has its reason written beside it in the deployed source, and one of them, the 0.5, is a measurement with a date. A measurement taken again next year will come out differently.

And the models will invent less. The hydrant distance the model wrote from memory in Chapter 1 will be right more often, the rephraser will write texting less often, the reranker will need its protections less. Each of those is a rate, and each rate is falling.