The verified eval marketplace
A marketplace of AI-agent eval packs — where nothing gets listed until it passes the authority. Every pack carries a grade, a discriminating-power benchmark, and a public audit trail.
All-access pass
Unlimited access to all verified packs and unlimited runs — plus every new pack the day it clears the authority. Best for teams gating more than two evals in CI.
Verifies a contract-QA agent cites the actual clause it answers from, and refuses to answer questions the contract doesn't cover.
Checks that a policy/claims assistant answers from the exact clause it cites — deductibles, limits, exclusions — and abstains on claim-specific questions it has no evidence for, instead of inventing coverage.
Verifies a benefits-eligibility assistant grounds answers in the actual program rule cited — thresholds, deadlines, required documents — and declines case-specific questions it can't evidence.
Verifies an employee-handbook assistant cites the actual policy section it answers from — leave, notice, benefits — and refuses personal HR questions it has no evidence for.
Verifies a shopping-support bot answers returns, shipping, and warranty questions from the actual policy cited, and declines order-specific questions (where's my package, why was I declined) it has no evidence for.
Verifies function/tool selection and argument correctness, and checks the agent asks for clarification instead of guessing when a request is ambiguous.
Checks that an accounting-policy assistant grounds answers in the actual policy note cited, and doesn't fabricate figures for uncovered questions.
A control cell: general-reference Q&A used to baseline groundedness and abstention scoring before applying the harness to specialized domains.
Checks that an infosec-policy assistant grounds answers in the actual control cited — rotation windows, encryption standards, incident SLAs — and abstains on live operational questions it can't answer from policy.
Evidence-based checklist for high-risk obligations: logging, human oversight, transparency, data governance, and conformity assessment. Anchored to the Act, not to opinion.
Verifies a support bot answers policy questions from the actual help-doc cited, and declines account-specific questions it has no evidence for.
Checks that a developer-docs assistant answers from the actual API reference cited — rate limits, status codes, auth — and abstains on environment-specific questions it can't answer from the docs.
Catches confident fabrication with fake citations. Scores groundedness, citation accuracy, and whether the agent abstains when evidence is missing.
Adversarial web content that tries to make a computer-use agent exfiltrate data or take destructive actions. The test set is the attack, not a Q&A.
An early-draft submission that hasn't cleared verification yet — held back rather than approved for sale. The report below shows exactly which checks it didn't pass.
For pack authors
Turn a battle-tested eval pack into revenue. Submit once — if it clears the six-axis verification, it's listed on the marketplace with a grade buyers trust. You keep the majority of every sale.
You price it; we take a flat marketplace fee. Payouts on every sale.
Buyers see your grade, benchmark and audit trail — not just a description.
Don't pass? You get the exact failing checks back — fix and resubmit free.