Abstract

Computations, model outputs, tool invocations and delegated requests do not, by themselves, carry authority to cause real-world effects. Yet in AI agents, distributed services, cloud platforms, financial systems and connected devices, an operation that was approved at one layer can change before it takes effect. It may be rewritten by a tool, split across services, delayed, retried, reassigned to different hardware or sent through a path nobody was watching. A credential that proves who is calling does not show which operation is about to happen, and a log written afterward cannot undo an irreversible effect.

Based on the Indian Provisional Filing filed 23 October 2026, which is built on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026.

This paper describes a protected execution-finality architecture for that problem. A proposed operation is held in a non-effective state. Evidence about the originating authority and the proposed operation is collected and independently validated, and in some embodiments the evidence gathered when the operation is proposed and when it is executed is cross-committed. A scoped, non-bearer authorization is issued. At a controlled effect boundary, an enforcement point obtains or reconstructs the operation that is actually about to execute and checks its effect-relevant parameters against current authority, validation evidence, policy and revocation generations and its own protected local state. If the checks pass, the operation is released. If authority is missing, mismatched, stale or already consumed, the operation is withheld or sent to qualified recovery. Implementations include network transmission, persistent-state commits, credential activation, financial operations, cloud-resource changes, information disclosure and device actuation.

The disclosure sets out forty non-limiting embodiments. They cover continuity of an approved effect across software layers, source-neutral effect semantics, detection of changes between observation and effect, enforcement across every effect-capable path in a declared domain, authority that lives at the resource and is never handed to the requester, bounded streaming and lease-based authorization, protected activation of policy, risk-scored routing and multi-tier independent human or machine review, cumulative risk accounting, independent runtime witnesses, coverage across delegated services and devices, crash-consistent journaling with reconciliation of unknown outcomes, and hardware-level control of device paths including DMA, command queues and accelerator output. Optional algorithmic-logic verification can add assurance about which software or model is running, but it does not replace verification of each operation at the effect boundary.

The scope is stated honestly. The embodiments are combined only where technically compatible and supported by the applicable disclosure. The paper makes no measured performance or security-improvement claim and no claim of protection for effect paths outside the enforcement boundary. It is not a standard, a certification, a regulatory compliance statement or a legal opinion, and it is not an endorsement of any company or product.

Where to read the full PCT filing: search “Sangam Das” on WIPO PATENTSCOPE (https://patentscope.wipo.int/); search PCT/IB2026/055615 (pctib202655615); or read the PCT file directly in TD Commons at https://www.tdcommons.org/dpubs_series/12542/.

Disclaimer: This abstract and the drafting were prepared with AI assistance and may contain mistakes. Readers must cross-check them against the original public disclosures and the patent filing copy attached to this record before drawing any conclusion.

License: CC BY-NC 4.0.

Keywords

AI Safety, AI Governance, Trustworthy AI, AI Governance Framework, AI Policy Enforcement, Machine-Enforceable AI Policy, Governance by Design, AI Risk Management, AI Accountability, Responsible AI, AI Agent Security, Agentic AI Security, Autonomous AI Agents, Autonomous Systems Security, Runtime Authorization, Operation-Bound Authorization, Consequence-Bound Authorization, Protected Execution-Finality, Execution Control, Effectuation-Boundary Enforcement, Policy Enforcement Point (PEP), Policy Decision Point (PDP), Zero Trust Architecture, Reference Monitor, Complete Mediation, Non-Bypassable Execution Control, Resource-Side Authorization, Non-Bearer Authorization, Sink-Local Authorization, Tokenless Authorization, Capability-Based Security, Least Privilege, Time-of-Check/Time-of-Use (TOCTOU) Prevention, Confused Deputy Prevention, Anti-Replay Protection, Single-Use Authorization, Fail-Closed Enforcement, Revocation Checking, Policy Versioning, Cross-Layer Security, Effect Continuity, Effect-Path Coverage, Operation Reconstruction, Independent Validation, Cross-Committed Evidence, Evidence-Carrying Receipts, Tamper-Evident Audit Evidence, Streaming Authorization, Lease-Based Authorization, Risk-Based Decision Routing, Risk-Weighted Autonomy, Cumulative Risk Accounting, Multi-Tier Review, Independent Human Review, Human Oversight of AI, Human-in-the-Loop Control, Independent Runtime Witness, Runtime Observation, Delegation Chain Security, Delegated Service Security, Crash-Consistent Journaling, Indeterminate Outcome Reconciliation, Failure Recovery, Hardware-Enforced Security, DMA Protection, Device Command Queue Security, Accelerator Output Control, GPU Security, Device Reassignment Security, Firmware Attestation, Hardware Root of Trust, Remote Attestation, Confidential Computing, Trusted Execution Environment (TEE), Cloud Resource Control, Financial Operations Security, Credential Activation Security, Information Disclosure Control, Industrial Automation Security, Cyber-Physical Systems Security, Critical Infrastructure Security, Software Integrity Verification, Digital Sovereignty, EU AI Act, NIS2, Cyber Resilience Act (CRA), IETF Internet-Draft, Patent-Pending Technical Disclosure, Technical Disclosure Commons, Candidate Act, Non-Effective State, Finality Sink, Actual Pending Operation, Protected Enforcement Domain, Algorithmic Logic Fingerprint (ALF)

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Share

COinS