Scores you can defend, by meaning — not by wording.
Rule checks first, meaning-based comparison second, calibrated judges third. Every answer is scored on accuracy, completeness, relevance, clarity, consistency, grounding, citations, policy, safety, task completion and trajectory — each with its reason — and judges count only where they agree with your experts.
Evaluation · included from the Team plan
What it does
- Rubrics with written definitions, weights, critical dimensions and examples, versioned
- Claim-by-claim verification against approved sources: supported, unsupported or contradicted
- Five citation checks: resolves, approved, current, supports the claim, permitted for this user
- Policy packs: disclosure (Texas TRAIGA), personal data, advice, referral, toxicity and your own rules
- Pinned judge models from a different family than the agent, calibrated against SME labels (agreement, Cohen’s kappa, false-pass rate)
- Quality gates by risk tier, approved before they take effect
Works with
- Runs & regression
Every release compared with the last approved one.
See evaluation with your data
We’ll connect a test environment of your agent and a few of your policy documents for the demo.
Request a demo