DSPy: MIPROv2 versus GEPA

DSPy is the one exercise in this course that needs a model API rather than your subscription, so the design starts there: one .env picks an OpenAI-compatible endpoint, and every model in the project is built from it. After that the job is fairness. Both optimisers start from the same signature, see the same 20 train and 10 val packs and are scored by the same metric as your claude -p loop; then all five prompts are run on the same 10 test packs with the same task model. The decision that matters most is estimating the cost from the optimisers' own budget rules before running, which shows where the money goes (reflection, not the task model) and lets you stop if the number surprises you.