Proof for what your AI says.
PRAMAN checks a generated output against the evidence it should rest on, and returns, per claim, a calibrated verdict, a distribution-free bound on the rate of auto-approving an ungrounded claim, an accept / escalate / reject decision, and a regulator-ready audit record. It bounds a rate, it is not a per-item certificate; it keeps the human on the decisions that matter.
then run a verification.
A bound you can defend, not a vibe score.
Decompose
Split the output into atomic, independently checkable claims.
Score & calibrate
A small CPU model scores each claim against the evidence; scores are calibrated into real probabilities.
Conformal control
Conformal risk control picks the threshold so the missed-approval rate is provably ≤ α.
Decide & log
Accept the safe majority, escalate the uncertain, and write a content-hashed audit record.
Stated truthfully: PRAMAN gives provably right-sized review and a defensible audit trail. The guarantee is marginal, not a per-item certificate; it is faithfulness to the supplied evidence, not truth in the world; and it does not remove the human on catastrophic or irreversible decisions. This managed-cloud page is a demo; the product runs air-gapped on the operator's own hardware.