Abstract
Patent pending. Publication Number WO/2026/115520; Publication Date 04.06.2026; International Application No. PCT/IB2026/053385; International Filing Date 07.04.2026. Published title: “Hardware-Enforced Cryptographic Execution-Time Governance Infrastructure for Artificial Intelligence Machines, Satellite, Digital Currencies, and Autonomous Systems Using Capability Withholding, Algorithmic Logic Fingerprinting, and Fail-Closed Enforcement at Irreversible Execution Boundaries” (https://patentscope.wipo.int/search/en/detail.jsf?docId=WO2026115520). It claims priority from Indian provisional applications including 202531130168 (22 December 2025), 202631007467 (26 January 2026) and 202631018571 (18 February 2026). The later application PCT/IB2026/055615 (filed 4 June 2026, publicly disclosed 16 July 2026) claims priority from this application. Inventor: Sangam Das. Searchable on WIPO PATENTSCOPE.
The AI industry is building strong ways to prove which file a model is and which software a server runs:
- Google, NVIDIA and HiddenLayer led the OpenSSF Model Signing standard.
- NVIDIA signs every model in its NGC catalogue.
- Apple’s Private Cloud Compute publishes a tamper-evident log of the software its servers run.
These steps show where a model came from and what code is running. They do not, by themselves, show what the model actually decides. A model can be fine-tuned, quantised, wrapped in a prompt or proxy layer, or replaced while still presenting a valid signature or name. An AI agent can act on its output before anyone notices.
This application describes Algorithmic Logic Fingerprints (ALF): compact cryptographic fingerprints of an algorithm’s decision behaviour, generated without access to source code, model weights or training data.
- For machine-learning models, the fingerprint is built by sampling decision boundaries.
- For traditional algorithms, it is built from normalised control-flow and decision structure.
- The highest-assurance form probes the model with a sealed set of test inputs that the computing system cannot see in advance.
- It then requires four views to match together: the decision-boundary profile, the monotonicity and consistency profile, the constraint-activation profile and the constraint-topology hash. This means wrapping a different model to imitate an approved one is structurally defeated.
Approved fingerprints are kept in a registry. At run time, inside a trusted execution environment, the live fingerprint is compared with the approved one after computation finishes and before the output takes effect. If the logic has changed, approval is automatically invalidated and the output is withheld (fail-closed). Each decision produces a tamper-evident, ledger-anchored validation receipt. Applications include AI systems, satellites, digital currencies and autonomous systems.
Industry relevance: ALF could complement model signing on NVIDIA NGC and Google Kaggle, verifiable AI clouds such as Apple Private Cloud Compute, confidential computing from Google Cloud, Microsoft Azure, AWS, NVIDIA, Intel and AMD, and agent platforms from Google, Microsoft and Amazon Web Services. It does this by adding verification of decision behaviour at the moment an output takes effect. Potential benefits extend to payments (Visa, Mastercard, PayPal, Ant Group), to vehicles under UN Regulation 156, to robots and satellites, and to Chinese AI developers including Alibaba, Baidu, ByteDance, Tencent, Huawei and Zhipu AI. ALF also supports compliance with the EU AI Act’s substantial-modification rule and China’s generative AI filing regime.
Disclaimer: Company and product names, including in the title, indicate technical relevance only. This abstract and the industry-relevance material are an AI-generated summary based on the original application and public sources. They describe potential benefits only. They are not an endorsement, partnership, sponsorship, validation or evaluation by any company named, and they make no allegation that any company uses, needs or infringes this invention. All company and product names are trademarks of their respective owners. Readers must read the original application on WIPO PATENTSCOPE (WO/2026/115520) before drawing any conclusion. The inventor is open to strategic collaboration or acquihire discussions.
Keywords
AI governance, algorithmic logic fingerprint, behavioural fingerprinting, model identity, model integrity, model signing, OpenSSF Model Signing, Sigstore, verifiable transparency, confidential computing, decision-boundary sampling, sealed-probe verification, wrapper resistance, approved model registry, automatic approval invalidation, trusted execution environment, execution-finality gate, fail-closed, validation receipt, EU AI Act substantial modification, China generative AI filing, UN R156, Google, NVIDIA NGC, Apple Private Cloud Compute, AI agents, agentic payments
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Recommended Citation
Das, Sangam Kumar, "Signed Is Not the Same as Behaving the Same: Algorithmic Logic Fingerprints That Verify What an AI Model Decides, Not Just Which File It Is, Relevant to Google, NVIDIA and Apple AI Trust Roadmaps (WO/2026/115520)", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12907