Abstract
Verifiable Third-Party Decision Binding at the Execution Boundary — With Alignment to Frontier AI Guardrails, Agent Protocols and AI-Governance Frameworks
AI agents and high-consequence systems routinely act on decisions made by someone else: a safety classifier or guardrail model, a policy engine, a fraud or sanctions service, a human approver, or a remote-attestation verifier. The component that actually sends the message, moves the money, deploys the change or calls the tool often cannot tell a real external decision from a claim that "the check passed". A compromised or poorly designed intermediary can skip the check and invent an ALLOW, replay an old decision, attach approval from a different action, hide a DENY, present only one of several required approvals, or keep the evidence for audit while the effect never depends on it. OAuth introspection, transaction tokens, HTTP message signatures, RATS attestation results, SCITT receipts and signed permits each solve part of this. This disclosure defines the remaining requirement as an effectuation-time invariant: a deployment declares which external authorities must decide, including mandatory, alternative, threshold, conditional and deny-overrides rules; each decision must be shown by independently verifiable evidence or an authenticated live query bound to the exact action, the real decider, the current policy and revocation state, freshness, the intended sink and permitted use; and the action stays non-effective until a finality sink verifies the complete set and commits atomically. Otherwise the system provides auditability, not prevention. The disclosure includes a formal model, pseudocode, attack analysis and test vectors. Alignment is discussed with publicly described guardrail, approval and tool-call mechanisms at OpenAI, Anthropic, Google, Meta, Microsoft and AWS, and with IETF, MCP, A2A and AI-governance frameworks; this alignment material is AI-generated and is not industry validation, so readers must cross-check it against the original patent filings. This disclosure is based on PCT/IB2026/055615 (the "Mothership" application), filed 4 June 2026 and disclosed 16 July 2026, which claims priority from PCT/IB2026/053385, filed 7 April 2026 and disclosed 4 June 2026. The author is open to acquihire or strategic collaboration. IETF Datatracker: https://datatracker.ietf.org/doc/draft-das-third-party-decision-binding/
KEYWORDS
external decision verification, guardrail decision binding, AI agent approval enforcement, human-in-the-loop approval binding, policy decision point evidence, fraud and sanctions check binding, exact-action binding, required authority set, threshold and deny-overrides approval, decision replay prevention, stale decision revocation, evidence before effect, finality sink verification, OAuth token introspection, SCITT transparency receipt, RATS attestation result, auditability versus prevention, fail-closed enforcement
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Recommended Citation
Das, Sangam Kumar, "Trust Me, I Checked": Why AI Agents and High-Consequence Systems Must Prove the Guardrail, Policy Engine or Approver Really Said Yes Before Anything Happens", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12812