Your agent presents signed authority to a bank, supplier or counterparty. Their system verifies it offline against your public key before acting. Yours independently verifies the authority presented by theirs. No prior integration. No call to NOMOS. No model in the decision.
loading the verifier…
NOMOS reconciles policy, behavior and outcomes into a sealed, machine-executable authority artifact, then verifies every decision against it before it runs.
Written policy is a claim. Behavior is evidence. Outcomes are proof. NOMOS reconciles all three and seals the result as something your systems can call.
Your policy, reconciled with how your teams actually decide, becomes sealed, versioned rules: conditions, outcomes, priorities, a default, escalations and a data contract. Your key signs it. Start from a template or fork one in Studio against your own documents.
An honest agent that was misled looks exactly like a compromised one. So authority travels inside the act, and the system the action would change checks it. An agent that holds no authority cannot be a confused deputy.
No term for who asks. The verdict is AUTHORIZED, DENIED, ESCALATED or INCOMPLETE.
Nor does the agent choose which law applies. The system that would change works out every authority that governs the action, from registrars it pins, and commits only when each one allows it. Part III ↗
The exact action, the sealed authority, testimony from certified witnesses, approvals bound to this exact act, an audience, a nonce and a short window. The agent holds no key that matters.
Nine checks, stopping at the first failure. Facts come from witnesses or its own records, never from the agent. A missing fact is undecided, never false.
On AUTHORIZED it applies exactly the action, records the nonce and re-reads its own facts in one step. Anything else commits nothing.
Point your agent at the authority. Every decision returns authorized, denied or escalated, with the rule that decided it and a hash you can hand to an auditor.
No API key, no signup — nomos.can() queries any published Exchange authority directly and hands back a verdict with a permanent, independently verifiable transcript URL.
Reality keeps moving. NOMOS tracks outcomes after every verified decision and flags when behavior starts drifting from the sealed artifact again — before it becomes an audit finding.
Agent security starts with permissions: who the agent is, and what it may do. But when injected text misleads an agent, its identity, credentials and operator stay the same. An honest agent misled by injected text and a compromised agent are indistinguishable at the moment they act. So authority has to attach to the action, not the agent.
Permissions decide what an agent may do.
NOMOS decides what is allowed to change.
Before anything commits, the system being changed checks the action against your institution's signed rules. If it doesn't pass, nothing changes.
Start with one call: nomos.can() checks an action against signed rules. You state the facts, and it answers. Add an Act when the answer has to bind. Read the paper ↗
One high-stakes decision, usually the one an auditor or a regulator already asked about. Then the next, on the same authority.
Seal your first artifact free. Pay when you put it in production.
Straight answers about how NOMOS works, what it stores, and what it doesn't claim to do.
accept() works out every authority that governs an action, from registrars you pin, and commits only when the Act carries each one.
OIDC and SAML 2.0, verified domains, and require-SSO with an owner break-glass.
Seal and sign with keys held in AWS KMS, Google Cloud KMS, Azure Key Vault or your HSM.
An approver signs with a passkey bound to the exact Act. A touch alone is not an approval.