Corsac exists to make agent approval measurable and defensible.

What Corsac does

Corsac points at an AI product and generates a graded eval library for it: discover what the product actually exposes, map that surface into eval areas, draft scenario-based evals grounded in the evidence, and grade every case with an LLM judge against an expected-behavior rubric. Only passing evals graduate to the published library — a pass-only discipline that makes each pack a benchmark a team can defend, not a generic score.

We call the result opinionated evals: suites that encode a specific, defensible view of how a particular product should behave — concrete failure modes, severity-tagged cases, and a graded pass bar. The public library covers hundreds of enterprise AI products built this way, and the benchmark generator runs the same pipeline live against any product URL. Where human judgment is the bar, managed expert review adds domain reviewers on top of the automated grading.