Microsoft shipped an open-source framework this week that converts plain-text behavior descriptions into AI evaluation tests.
The tool is called Adaptive Spec-driven Scoring for Evaluation and Regression Testing, a name that covers more syllables than most machine learning papers. Released Tuesday, it gives developers a way to describe how a model should behave in natural language, then generates evaluations from those descriptions. The framework includes a regression testing component, designed to catch when a model update breaks behavior that previously worked.
Building reliable AI evaluations has historically required dedicated engineering work: writing test harnesses, defining scoring rubrics, and running them at scale. If text-based spec generation actually reduces that friction, it could make eval coverage accessible to teams without a dedicated ML research function. Regression testing in particular is underserved - models are updated constantly, and most teams have no systematic way to detect when a capability quietly degrades.
Whether a natural-language spec can be precise enough to catch the failure modes that actually matter is the harder question. Vague descriptions produce vague tests, and vague tests produce false confidence.