Abstract

Relevant to Microsoft, Google, AWS, OpenAI, Anthropic, NVIDIA, Meta, Apple, Alibaba, Baidu, Tencent, Huawei, SAP and Mistral AI

Most AI safety controls judge what an output says, through content filters, moderation classifiers and principle-based training. Others prove where code ran, or record what happened for later analysis. None of these, on its own, answers a different question before an output takes effect: did this particular AI run behave the way an approved run should?

Based on Indian Provisional Filing: 202631046689, filed 11 April 2026, and disclosed with the Mothership (PCT/IB2026/055615) on 16 July 2026.

This disclosure answers that question at runtime:

  • Signals.While an AI system processes a request, predefined instrumentation points collect signals about how that run is actually executing:
    • which policy path fired and whether a safety trigger tripped;
    • how the output was classified;
    • which tools and retrieval steps ran;
    • which model version was used;
    • the workflow state, purpose and risk tier.
  • Descriptor. These signals are condensed into a small, bounded behavioural descriptor of the run, without reconstructing the model’s internal reasoning.
  • Release check. Before the output is shown, returned, sent, stored or used to act, the descriptor is checked against behaviour classes approved before the run began. The output is released only on a match. Otherwise it is denied, constrained, transformed, redirected or escalated.

The decision depends on how the AI behaved, not on its content, so it complements content filters rather than replacing them. The disclosure also covers:

  • requiring both an approved behaviour class and an approved purpose class;
  • fail-closed handling when evidence is missing or tampered with;
  • enforcement in a wrapper or orchestration layer around third-party models, with no model changes;
  • per-unit control of streamed output;
  • privacy-preserving validation records for audit.

Industry relevance (AI-generated). This section was prepared with AI assistance and shows potential fit and collaboration opportunities only. The platforms below already produce the signals this architecture uses: guardrail results, tool-call traces and model versions. The disclosure describes turning those signals into a release condition.

  • United States:
    • Microsoft Azure AI Foundry, Content Safety and Copilot Studio;
    • Google Vertex AI, Model Armor and Agent Gateway;
    • AWS Bedrock Guardrails and AgentCore;
    • the OpenAI Agents SDK, with its guardrails and tracing;
    • Anthropic’s Claude agents and MCP;
    • NVIDIA NeMo Guardrails and NIM;
    • Meta’s Llama Stack and LlamaFirewall;
    • Apple’s Foundation Models framework;
    • enterprise platforms such as Salesforce, ServiceNow, IBM watsonx, Databricks and Snowflake;
    • observability vendors such as Datadog and LangSmith.
  • China:
    • Alibaba Cloud Model Studio and Qwen;
    • Baidu Qianfan and ERNIE;
    • Tencent Hunyuan and WeChat agents;
    • ByteDance Doubao, Coze and Volcano Engine;
    • Huawei ModelArts and Pangu;
    • Zhipu AI, DeepSeek and Moonshot AI.
  • Europe:
    • SAP Joule and Siemens Industrial Copilot;
    • Mistral AI and Aleph Alpha;
    • Deutsche Telekom, Orange, Vodafone, Ericsson and Nokia;
    • European banks and insurers such as Allianz, AXA, BNP Paribas and ING.
  • Elsewhere:
    • Samsung and Qualcomm on-device AI;
    • Indian enterprise and telecom providers such as TCS, Infosys, Reliance Jio and NPCI.

Latency and legacy. The check reuses signals the platform already produces, runs once per released response or streamed unit rather than inside the model, and can sit in an existing gateway or orchestration layer without changing the model.

Filing details. Indian provisional patent application no. 202631046689, filed 11 April 2026, available on WIPO PATENTSCOPE among the priority documents of the Mothership application PCT/IB2026/055615 (WO 2026/150382), filed 4 June 2026 and publicly disclosed 16 July 2026. Inventor: Sangam Das ([email protected]). The inventor is open to acquihire or strategic collaboration.

Disclaimer. The invention is the work of its human inventor. This title, abstract and the accompanying summary, including all company and product references, were prepared with AI assistance. Readers must cross-check every statement against the original filed documents on WIPO PATENTSCOPE and official company sources, and draw their own conclusions. Naming a company or product shows topical relevance only. It is not an endorsement, a validation, a criticism or a statement of infringement, and no company named is affiliated with the inventor. The application is pending, and nothing here is legal advice. Licence: CC BY-NC 4.0.

Keywords: AI safety; runtime behavioural descriptor; execution-path governance; output gating; release control; approved behaviour class; purpose binding; guardrails; content safety; AI runtime security; tracing; observability; OpenTelemetry; tool-call trace; model version pinning; fail-closed; streaming output control; wrapper governance; orchestration layer; third-party foundation models; agentic AI; RAG; validation record; privacy-preserving audit; EU AI Act; NIST AI RMF; ISO/IEC 42001; Azure AI Foundry; Vertex AI; Model Armor; Bedrock Guardrails; AgentCore; Agents SDK; Claude; NeMo Guardrails; Llama Stack; Foundation Models framework; Qwen; ERNIE; Hunyuan; Doubao; Pangu; SAP Joule; Mistral AI

Creative Commons License

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

Share

COinS