Evaluation Regression

A versioned comparison between a candidate configuration and a baseline to detect losses in quality, safety, cost, or performance. The change may affect the model, prompt, retriever, tool, policy, or data. The test needs a fixed protocol and population; otherwise, a dataset shift may look like a system regression.

Global averages hide localized losses. Analysis should report case- and segment-level differences, uncertainty, and operational severity. Real incidents feed the regression set, while a quality gate decides which losses block progress and which require review.