Toward Algorithmic Disarmament: A Socio-Technical Framework for AI Non-Proliferation and Legal Verification

The rapid penetration of foundation models into strategic decision-making and Lethal Autonomous Weapons Systems (LAWS) marks a paradigm shift, rendering the traditional kinetic weapon-centric disarmament frameworks (e.g., the Treaty on the Non-Proliferation of Nuclear Weapons) structurally ineffective in the digital age.This paper proposes a comprehensive Algorithmic Disarmament framework, advocating for a shift in regulatory granularity from physical hardware to model weights and training provenance. We construct a quantitative risk function: $R = \alpha \cdot \log_{10}(C) + \beta \cdot D_{\text{sens}} + \gamma \cdot A_{\text{ind}}$, which classifies high-risk AI assets into legal tiers based on computational scale ($C$), data sensitivity ($D_{\text{sens}}$), and autonomy index ($A_{\text{ind}}$), thereby defining the legal boundaries of algorithmic weapons. To resolve the inherent tension between state sovereign secrets and international regulation, this paper demonstrates the technical feasibility of non-intrusive auditing using Zero-Knowledge Proofs (ZKP) and cryptographic watermarking, enabling international monitoring bodies to verify whether models comply with humanitarian constraints without infringing on intellectual property rights or national security classified information. By evolving the Martens Clause into a socio-technical norm for the algorithmic era, this research conceptualizes the operational architecture of an International Artificial Intelligence Oversight Organization (IAIO), providing a robust legal and technical roadmap for curbing the unregulated AI arms race and safeguarding global peace.