Radoslav Y. Radoslavov — Attorney, Legal Engineering Practice
Executive Summary
The Zero-Knowledge Audit Protocol (ZKAP) proposes a structural approach to the information asymmetry between regulators and AI providers. By transforming selected legal obligations under the EU AI Act (Regulation 2024/1689) and the NIS2 Directive (Directive 2022/2555) into cryptographic polynomial constraints, ZKAP aims to enable compliance verification without access to proprietary models, training data, or personal information processed at inference. The protocol is designed for integration at the software layer today and at the hardware layer over a longer implementation horizon.
Addressing the Scale Problem in AI Oversight
Traditional oversight methods were developed for systems whose behaviour could be reconstructed through inspection. High-complexity AI systems processing continuous inference at scale exceed what human-led manual certification can practically verify — particularly when direct inspection of model internals would itself conflict with trade secret protection (Directive 2016/943) and data minimisation obligations (GDPR Articles 5, 9).
ZKAP addresses this structural limitation through three mechanisms:
• Execution-time attestation. Implementation at the software or hardware level ensures that each inference event generates a cryptographically verifiable record that cannot be retroactively modified.
• Verifiable computation. The approach moves from declarative compliance statements to mathematical proofs of specific properties at the moment of computation.
• Zero-Knowledge Proofs. Regulators and notified bodies receive compact proofs (typically under 2 MB) that confirm satisfaction of formalised constraints in milliseconds, while the model internals, input data, and intermediate state remain inaccessible to the verifier.
Institutional Positioning
• Institutional Licensing. We are open to discussing royalty-free licensing arrangements with qualifying public institutions and regulatory bodies on a case-by-case basis, subject to appropriate framework agreements.
• Formalisation of Legal Norms. By translating quantifiable regulatory requirements into polynomial constraints, the protocol preserves human regulatory judgement at the formalisation stage while enabling mathematical verification at the enforcement stage. Qualitative legal concepts — reasonableness, proportionality, legitimate interest — remain outside this scope and in the domain of human interpretation.
• Applications in Public Administration. Potential applications include verifiable algorithmic decision-making in public administration, with mathematical guarantees against demographic bias where such bias can be formally specified.
Scope and Honest Limitations
No working prototype currently exists. The architecture is specified in accompanying technical documentation. Independent peer review has not yet been conducted. Approximately 70% of the cryptographic stack required for a software reference implementation exists as production-grade open-source libraries; the remaining engineering work is estimated at approximately seven months. These limitations are stated openly because the proposal at this stage invites regulatory dialogue, not premature deployment claims.
Engagement
The August 2026 deadline for Article 5 of the AI Act, and the August 2027 deadline for Annex III high-risk systems, mark the institutional timeline within which these questions require answers. For dialogue: advanced-consulting.london · radoslavov.bg
Patent applications covering the ZKAP architecture have been filed; international filings are in preparation.