Dspy Gepa Optimizer

If your metric returns rich critiques (not just a float), GEPA beats MIPROv2 on complex tasks with fewer rollouts. The catch: you need a metric that returns dspy.Prediction(score=float, feedback=str), a strong reflection LM at temperature 1.0, and separate train/val splits where you actually maximize training data instead of the usual validation-heavy split. Use auto="medium" for everyday work, auto="heavy" before shipping. The feedback quality is the load-bearing part: vague critiques waste rollouts, specific ones about why it failed and what good looks like are what the reflection LM acts on.