Abstract

Causal semantic handoff verification for multi-agent AI. A shared, jurisdiction-neutral reliability layer for U.S., Chinese and global AI ecosystems (Indian Provisional Filing 202631134241)

ABSTRACT

The United States and China are building the two largest ecosystems of autonomous AI agents. Models, agent frameworks and cloud platforms from both sides increasingly call one another, directly or through third-country services, to run banking, cloud operations, telecom and energy workflows. When such a chain fails, for example when a spending limit disappears or a web page is treated as a command, a hard question follows. Where did the meaning of the human’s instruction change, and whose component changed it? Today’s message authentication, tracing, guardrails and tool authorization show that a message arrived and that an action was permitted. They cannot show where meaning was lost. In an era of export controls, sovereignty concerns and mutual distrust, that gap makes every multi-agent incident easy to dispute and easy to misread as intent.

Based on Indian Provisional Filing 202631134241, filed 8 November 2026. This record contains that filing only. The Mothership application PCT/IB2026/055615 (filed 4 June 2026, publicly disclosed 16 July 2026, WO 2026/150382 A2) is a separate filing and is not attached.

The disclosed method is a diagnostic layer for multi-agent AI safety. It represents each piece of content with its source and trust role, so a fact from a low-trust source is distinguished from an instruction from an authorized principal. It runs matched counterfactual tests, changing a fact and an instruction-like statement separately, to see what actually moves an agent’s behavior. It selects test cases adaptively, according to which conditions an action depends on. It localizes the earliest hop at which an action-relevant discrepancy becomes observable, and it reports uncertainty when the evidence is not enough. It runs through protocol-independent adapters, so no side has to change its protocol, its model or its enterprise software.

Why this matters for U.S.–China AI trust:

  • Evidence that neither side has to take on trust. A finding names a hop and an observable cause, from tests on behavior. It does not require access to either side’s model weights, training data or source code.
  • Interoperability without convergence. Platforms such as Google ADK/A2A, Microsoft Agent Framework, OpenAI Agents SDK, Claude Agent SDK/MCP, Alibaba AgentScope and Tencent Cloud ADP can keep their own protocols and policies while emitting the same kind of diagnostic finding.
  • Lower escalation risk. An unexplained AI-driven fault in critical infrastructure can be read as hostile. A located, evidence-backed fault gives governments and operators a factual basis to compare.
  • Sovereignty-compatible. Each jurisdiction and operator keeps its own rules of authority and its own data. The method checks whether an instruction’s meaning was preserved under those rules and does not impose a common rulebook.
  • Complement, not replacement. It sits beside guardrails, authorization, human oversight and protected execution controls, and adds evidence about where instruction meaning changes.

How it could sit in practice (potential integration, not existing implementations):

  • In ADK/A2A or Microsoft Agent Framework applications, an observer on delegated-task messages compares the originating instruction with the handoff version and flags a changed limit or condition.
  • In OpenAI Agents SDK or Claude Agent SDK/MCP applications, the check compares the originating instruction, the handoff representation and the proposed tool call, and flags a call that exceeds the original condition.
  • In AgentScope, Tencent ADP and Qwen-, DeepSeek-, Doubao-, Kimi- or GLM-based agents, the same adapter tests whether a lower-trust statement was treated as authority to change the task.

Performance, latency and diagnostic accuracy require independent experimental validation. No government position or adoption by any company is implied. The inventor is open to strategic collaboration, including acquihire: [email protected].

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

AI-generated notice: the industry-relevance and geopolitical text above was prepared with AI assistance. Readers must cross-check it against the filed Indian specification and the original PCT filing before drawing any conclusion. References to companies, products and jurisdictions illustrate potential technical relevance only. They are not an endorsement, validation, partnership, adoption or infringement statement. Licensed under CC BY-NC 4.0.

KEYWORDS

AI safety; U.S.–China AI; AI geopolitics; AI governance; agentic AI; multi-agent systems; agent-to-agent (A2A); Model Context Protocol (MCP); semantic drift; instruction fidelity; prompt injection; counterfactual testing; causal fault localization; AI observability; AI assurance; AI interoperability; AI export controls; digital sovereignty; critical infrastructure; AI risk management; NIST AI RMF; ISO/IEC 42001; EU AI Act; Indian patent application 202631134241

INDUSTRY RELEVANCE

AI-generated alignment, potential integration only. Each adapter sits beside the platform at a point the application already controls, reads only what the application may see, and checks whether each condition of the original instruction survived each handoff. It reports a finding, a handoff and an uncertainty record. It rejects and approves nothing, because the platform’s own authorization, guardrails and approval controls keep making those decisions.

  • Google (Agent Development Kit, Agent2Agent, Gemini Enterprise Agent Platform): a diagnostic extension on an ADK workflow-event interface or in A2A client/server middleware observes task submission, messages, task-state updates and returned artifacts. It compares each version with the originating conditions. If a supplier’s counteroffer became an “approved budget increase”, it reports the stage where that happened. The A2A protocol does not need redesign.
  • Microsoft (Agent Framework, Copilot Studio, Foundry): diagnostic middleware on the workflow and middleware layers captures handoff records and proposed tool arguments. It compares the original instruction, the communicated meaning and the proposed operation, and flags the handoff where a risk limit changed. It also checks whether earlier findings still hold after data updates. Existing orchestration, identity, governance and approval controls stay in place.
  • OpenAI (Agents SDK, handoffs, tracing): an adapter on handoff callbacks, input-transformation hooks, tool wrappers and permitted tracing processors compares the outgoing instruction, the incoming task representation and the proposed tool operation. Paired tests show whether an omitted condition changed behavior, and whether a retrieved document influenced a governing condition. It needs no access to private model reasoning.
  • Anthropic (Claude Agent SDK, Claude Managed Agents, Model Context Protocol): hooks in tool-result processing, delegation, message normalization and approved MCP interfaces separate a source’s fact-reporting role from any authority to change the instruction. Facts are held fixed while an instruction-like assertion is varied, using non-effectuating calls. It complements existing tool permissions, safety measures and access controls.
  • Alibaba (AgentScope, Qwen), Tencent Cloud ADP, and agents built on DeepSeek, Doubao, Kimi and GLM: the same adapter tests whether a lower-trust statement was treated as authority to change a budget or delivery condition, so cross-border chains can be examined with one shared method.

Readers must cross-check this against the filed Indian specification (202631134241) and the original PCT filing before drawing any conclusion. This alignment is not an endorsement, validation, partnership, adoption or infringement statement, and no company has reviewed or accepted it. Performance, latency and accuracy require independent validation.

Creative Commons License

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

Share

COinS