Eval

Eval answers the go/no-go question. Define a target (a model plus its system prompt), a test set of prompts, and the criteria that matter (accuracy, relevance, completeness, clarity, safety, or your own). Each case is answered by the target on the live relay, then scored by a judge model against every criterion.

Per-case pass/fail against a threshold rolls up to a pass rate and a READY / NOT READY production-readiness verdict, with per-criterion averages and a failures view. Judge scores are labelled as a model's opinion, not ground truth; a case that fails to generate is a failure, not a skipped row.

A Pro workspace gated behind Nyquest Pro, enforced frontend and backend. Export a portable Nyquest Eval Report.