“Send $10,000 to a vendor for invoice #4821”This is a real-world action. --- ## Scene 2 — Signal creation Instead of executing, the system converts the AI output into a structured signal: Action: PAYMENT_APPROVAL Amount: $10,000 Recipient: Vendor X Reason: Invoice #4821 --- ## Scene 3 — Governance evaluation The signal is sent to Parmana. Parmana evaluates: - signed policy rules - authorization constraints - execution boundaries --- ## Scene 4 — Authorization Decision Parmana returns:
APPROVEDOr:
REJECTEDOr:
REQUIRE_OVERRIDEThis decision is deterministic and reproducible. --- ## Scene 5 — Execution gating Only if APPROVED: The Execution Runtime executes the payment. If not: The action is blocked. --- ## Scene 6 — Cryptographic attestation Every decision produces a cryptographic proof: - policy version - signal hash - decision outcome - executionId - runtime identity This can be independently verified. --- ## Scene 7 — Audit trail Any external party can verify:
“Was this payment actually authorized?”Without accessing internal systems. --- ## Core message AI can propose actions. Parmana decides if they are allowed. Execution happens only after authorization. --- ## Why this demo matters This is not a simulation problem. This is a real-world execution safety problem: - payments - infrastructure changes - enterprise actions - autonomous AI agents --- ## Final takeaway
Parmana ensures AI systems cannot execute real-world actions without explicit authorization.