Guardian

GuardMeta Treasury is a proof-carrying authorization system for AI-assisted treasury workflows. It addresses a critical gap: an AI model can generate perfectly valid JSON while still proposing an unauthorized payment. Meta Muse Spark Open, served through Modal, is GuardMeta’s reasoning component. GuardMeta exposes six controlled evidence tools, backed in this prototype by deterministic demo records representing IAM, ledger, approval, beneficiary, policy, and idempotency services. Muse chooses the evidence calls and constructs an immutable DRY_RUN transaction proposal. GuardMeta then independently validates the trusted request and evaluates ten authorization policies, including approval binding, separation of duties, available balance, rolling daily exposure, beneficiary trust, currency controls, and replay protection. The policy layer uses deterministic checks and Z3 constraints to produce an auditable SAT or UNSAT certificate with named checks, counterexamples, a policy version, and a certificate digest. The deployed benchmark passed all three live scenarios: one authorized payment, one requester self-approval denial, and one rolling-exposure denial.