You deploy a Meta-Optimizer to rewrite system instructions based on recent failure logs. How do you mathematically guarantee the new prompt doesn't degrade performance on edge cases that used to work?
AI Engineer Problems › Evaluation & Testing
Evaluating Prompt Regressions
Medium · AI Engineering · Evaluation & Testing · Asked at Rocket Lawyer, Squaretrade, Quantifi Solutions