Abstract

From GPT, Claude and Gemini to Qwen, DeepSeek and Huawei 6G: protected execution finality, independent operation reconstruction, scoped authorization, AI-agent security, 6G infrastructure, enterprise data governance, digital payments and critical infrastructure.

Abstract

The U.S.–China AI race is no longer decided only by whose model scores higher. As GPT, Claude, Gemini and Grok on one side, and Qwen, DeepSeek, Doubao, Kimi and GLM on the other, gain the ability to call tools, delegate to other agents, move money, change cloud settings and reconfigure networks, the decisive question changes. It is no longer “How smart is the AI?” but “Who controls what actually happens at the moment of action?” An AI request can be correctly signed, authenticated and approved, and still turn into something different by the time it takes effect. It may be delegated, rewritten by a tool, delayed, offloaded to hardware or run under a changed policy, so the real operation is no longer the one that was authorized. This risk belongs to every country and every vendor, and no single ecosystem can solve it for the others.

Based on Indian Provisional Filing: 202631136555, filed 15 November 2026, and its Broad concept is disclosed with the Mothership (PCT/IB2026/055615) on 16 July 2026.

This disclosure explores an execution-time AI safety architecture that separates authority to compute, propose, prepare, simulate or request an operation from authority to produce its actual consequential effect. The proposed sequence is:

  1. Originating authority.
  2. Candidate Act.
  3. Non-Effective State.
  4. Actual Pending Operation reconstruction.
  5. Current authority and dependency validation.
  6. Scoped execution authorization.
  7. Protected Finality Sink.
  8. Actual consequence.
  9. Outcome reconciliation.

The 39 embodiments cover the following areas:

  • AI agents and agent-to-agent communication (A2A, MCP);
  • confidential computing and accelerators;
  • AI-native 6G and AI-RAN networks;
  • enterprise databases and ERP;
  • agentic payments and digital currency;
  • critical infrastructure.

The architecture is vendor-neutral. It is meant to work alongside existing authentication, guardrails, zero-trust controls, trusted computing and telecom standards, not replace them. Each jurisdiction and operator keeps its own rules about who may authorize what. It cannot claim non-bypassable control over endpoints that do not take part in its enforcement arrangements, and its security, latency and interoperability results need implementation-specific testing.

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

1. The Problem in Plain Words

Frontier AI now works as a chain: reasoning, planning, delegation, tool selection, authorization, execution, actual consequence. Security controls protect many of these steps. But the state seen at approval time can differ from the state present when the operation takes effect. An instruction approved under one identity, configuration or policy may run under another identity, resource state, destination or authority epoch. Four examples:

  • AI agents: An agent is permitted to modify one customer record. After delegation or tool transformation, the real operation touches a different customer or invokes extra privileged operations.
  • Digital payments: An agent may pay within a limit. Several individually allowed payments together exceed the approved exposure, or the final beneficiary differs from the authorized one.
  • 6G networks: A controller proposes a valid configuration change. By execution time, topology, shared resources or slice conditions have changed, and the effective change exceeds its authorized service-impact envelope.
  • Cloud and hardware: A software controller authorizes an operation, but a migrated accelerator, hardware-offloaded rule or stale execution domain produces an effect under a different authority or configuration.

In each case the earlier authentication may have been valid. The question concerns the actual operation at its effectuation boundary.

2. Core Technical Concepts

  • Authority to compute, propose, prepare, simulate or request an operation is separate from authority to produce its actual consequential effect.
  • A valid proposed action and a valid actual consequence are not automatically the same thing.
  • Candidate Act: the operation proposed by an AI agent, software system, human operator or autonomous controller.
  • Non-Effective State: the governed consequence cannot take effect until the authorization conditions are met. Unrelated computation and harmless preparation are not necessarily blocked.
  • Actual Pending Operation Reconstruction: determining the effective target, identity, parameters, route, material side effects, configuration and downstream consequences.
  • Independent Validation: evaluating originating authority, current policy, revocation, execution context, dependencies, evidence sufficiency and consequence-specific restrictions.
  • Scoped Execution Authorization: permission for the specific acceptable operation, using cryptographic capabilities, protected local state, hardware controls, trusted services or combinations.
  • Protected Finality Sink: the participating effectuation boundary that admits or withholds the governed consequence.
  • Outcome Reconciliation: separating request accepted, dispatched, committed, acknowledged and verified complete. Uncertain outcomes call for protected reconciliation, not uncontrolled retry.
  • Selective protected enforcement at consequential boundaries, not inspection of every token or instruction.
  • Intelligence determines what an AI system proposes; protected execution authority determines which consequential operations become real.

3. What the 39 Embodiments Address

  • AI agents and autonomous software: preserving authority through context compaction; detecting unauthorized changes in effective tool requests; validating aggregate results from cooperating agents; separating software generation from independently verified deployment; protected tool-execution interfaces; effects of differing model-provider protocols, hosted tools and developer-managed tools on where enforcement is possible.
  • Semiconductor security and confidential computing: accelerator authority transfer, trust-domain transitions, residual DMA and memory-access paths, numerical uncertainty, and whether an operation permitted in software is equivalent to the one actually installed in programmable hardware.
  • Telecom, AI-RAN and 6G: effective network-state authorization, cumulative connectivity changes, autonomous radio configuration, causal radio-service impairment, protected network-slice capacity lending. Complex validation may run outside the fastest loops, with bounded protected checks at configuration and authorization boundaries.
  • Enterprise AI, databases and cloud: effects from ERP systems, database triggers, stored programs, event consumers, data-policy transformations and analytical population activation, where a legitimate request produces additional effects downstream.
  • Digital payments and finance: cumulative agentic payment exposure, reservations, cross-border conversion and beneficiary outcomes, offline monetary delegation, post-compromise evidence, reconciliation of uncertain monetary effects.

4. How This Differs from Ordinary AI Governance

Identity management, tool permissions, monitoring, tracing, content restrictions, approval workflows and runtime checks are established and important. This architecture asks an additional question: does the actual effect-producing operation still satisfy the originating authority, current technical conditions, evidence requirements and permitted consequence? An AI tool gateway may confirm an agent may call an enterprise API. This architecture would additionally reconstruct the effective account mutation, identify material downstream effects, check current financial or data-governance restrictions, and require the participating effectuation interface to enforce the matching authorization. The distinction is not more logging or more signatures; it is the functional link among effective-operation reconstruction, current independent validation, scoped authority and actual enforcement.

5. Industry Relevance (AI-Generated, Hypothetical)

The mapping below is AI-generated. Readers must cross-check it against the original PCT filings before drawing any conclusion. It is not an endorsement, validation or infringement statement, and it was not checked against vendor documentation.

  • Anthropic (Claude Agent SDK, MCP): a hypothetical gate sits on the tool-execution path. It reconstructs the real tool call after delegation or context compaction, checks it against the originating authority and current policy, and withholds calls that exceed the authorized target or parameters.
  • OpenAI (Agents SDK, tool calling): a hypothetical gate sits at the tool wrapper or handoff step. It compares the effective operation after handoff with the one authorized and rejects a call whose record or beneficiary differs.
  • Google (agent platforms, A2A): a hypothetical gate sits at the agent-to-agent task boundary. It checks that a delegated task’s effective operation stays within the delegator’s authorization and revocation state at execution time.
  • Microsoft (agent infrastructure, Azure): a hypothetical gate sits at the workflow action step. It checks the effective cloud or enterprise change against current policy and blocks the effect if the policy epoch has changed.
  • Alibaba (AgentScope, Qwen), DeepSeek, Doubao, Kimi, GLM, Huawei: a hypothetical gate sits at the agent-orchestration tool interface and at AI-native network configuration. It reconstructs the real operation or slice change and withholds effects that exceed the authorized service-impact envelope.
  • Hardware, telecom and finance (NVIDIA, AMD, Intel, Arm, Ericsson, Nokia, Qualcomm, Visa, Mastercard, Swift, SAP, Siemens): hypothetical gates sit at accelerator reassignment, rule installation, configuration commit and payment submission. Each gate checks that the effective operation matches what was authorized and holds it if not.

6. Related Public Records

  • IETF Internet-Draft: draft-das-consequence-bound-execution-authority-00, an individual technical proposal and not an approved IETF standard.
  • Zenodo: https://doi.org/10.5281/zenodo.24274259
  • PCT/IB2026/055615 (the Mothership): international filing date 4 June 2026; published 16 July 2026 as WO 2026/150382 A2.
  • Indian Provisional Application 202631136555: filed 15 November 2026, based on the Mothership; 2,405 pages, 39 embodiments (1–38 and 12A). A shared foundation or terminology does not by itself make each specialized mechanism separately novel or patentable. Priority for later claims must be assessed against the actual filings.

7. Keywords for Patent Examiners

  • Core architecture: protected execution finality; execution-time authorization; consequence-bound execution authority; candidate act; non-effective state; actual pending operation; independent operation reconstruction; source-authority envelope; independent validator; scoped execution capability; finality sink; effectuation boundary; protected state commitment; hardware-rooted finality; authority continuity; effect-bound authorization; TOCTOU prevention.
  • Agentic AI and interoperability: agentic AI security; AI tool calling; agent-to-agent protocols; A2A; Model Context Protocol; MCP security; cross-protocol tool authorization; agent delegation; context compaction; multi-agent composition; AI agent governance; prompt injection defense; trusted tool execution; AI runtime security; zero trust AI.
  • Cloud and semiconductor: confidential computing; trusted execution environment; remote attestation; accelerator authority cutover; hardware offload equivalence; DPU security; SmartNIC security; DMA protection; protected key lifecycle; numerical error validation.
  • Telecom and 6G: AI-RAN; Open RAN; O-RAN; RAN Intelligent Controller; network slicing; autonomous networks; network configuration finality; causal radio impairment; connectivity trajectory validation; network slice capacity lending.
  • Enterprise and data governance: enterprise AI governance; database trigger security; source-commit certification; cross-engine policy equivalence; analytical cohort authorization; ERP transaction security; transitive execution effects.
  • Financial infrastructure: agentic payments; AI payment agents; CBDC security; offline digital cash; cross-border payments; financial exposure reservation; payment finality; transaction reconciliation; beneficiary verification; post-compromise monetary recovery.
  • Critical infrastructure and regulation: operational technology security; cyber-physical systems; autonomous industrial systems; EU AI Act; NIS2; DORA; Cyber Resilience Act.
  • Identifiers: PCT/IB2026/055615; WO 2026/150382 A2; Indian Provisional Application 202631136555; Sangam Kumar Das; draft-das-consequence-bound-execution-authority; Zenodo 24274259.

Suggested search strings

  1. (“candidate act” OR “proposed operation”) AND (“actual operation” OR “pending operation”) AND (reconstruct*) AND (authoriz*)
  2. (“non-effective state” OR “held pending”) AND (agent) AND (authorization) AND (effectuation OR “finality”)
  3. (“scoped” W/3 (capability OR authorization)) AND (operation) AND (“independent validator” OR “validation”)
  4. (“finality sink” OR “effectuation boundary” OR “enforcement point”) AND (agent OR autonomous)
  5. (TOCTOU OR “time-of-check”) AND (“execution-time” OR “effect time”) AND (agent OR AI)
  6. (“agent” W/5 “tool”) AND (delegat*) AND (authority) AND (“actual effect” OR consequence)
  7. (“hardware offload” OR accelerator) AND (equivalence) AND (authoriz*) AND (operation)
  8. (“network slice” OR “AI-RAN” OR “RAN intelligent controller”) AND (“autonomous” AND “authorization”) AND (effect*)
  9. (“agentic payment” OR “AI agent” W/3 payment) AND (cumulative OR exposure) AND (reservation OR reconciliation)
  10. (“uncertain outcome” OR “unknown outcome”) AND (reconcil*) AND (retry) AND (authority)

8. Notices

Technical source and AI assistance. The underlying inventions and patent filings are attributed to human inventor Sangam Kumar Das. This abstract, industry analysis, cross-references and keyword index were prepared with AI assistance. They may contain errors or omissions. Readers must verify patent identifiers, dates, claims, product descriptions and public references against the official filings and primary sources.

Neutrality and non-endorsement. References to the United States, China, Europe, their institutions and companies explain potential research relevance only. No adoption, endorsement, implementation, commercial relationship, copying, infringement or technological deficiency is alleged. No regulatory compliance, certification or deployment readiness is claimed.

Limits. The mechanisms depend on the trust assumptions, controlled effectuation routes and verification capabilities of each implementation. Novelty, inventive step, priority, performance, non-bypassability and patent scope must be judged against the actual filed documents, earlier disclosures and independent implementation evidence.

License: CC BY-NC 4.0

National Security, Critical Infrastructure Protection, Defense Cybersecurity, Sovereign AI, AI Safety, Trustworthy AI, U.S.–China AI Cooperation, Limited-Trust AI Collaboration, Agentic AI Security, Autonomous Systems Security, Zero Trust Architecture, Cross-Domain Security, Protected Execution-Finality, Non-Bypassable Execution Control, Runtime Authorization, Consequence-Bound Authorization, Policy Enforcement Point (PEP), Reference Monitor, Hardware Root of Trust, Trusted Execution Environment (TEE), Confidential Computing, Secure Multi-Agent Systems, Mission-Critical Systems, Cyber-Physical Systems Security, Operational Technology (OT) Security, 6G Network Security, Network Slicing Security, UAV and Autonomous Vehicle Security, Evidence-Carrying Receipts, Independent Effect Verification, Secure Command Authoriz

Creative Commons License

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

Share

COinS