ZKAP — Zero-Knowledge Audit Protocol: Reconciling AI Act Transparency, GDPR Privacy and Trade Secrets through Cryptography

Published: | Author: Radoslav Radoslavov, Attorney-at-Law

ZKAP — Zero-Knowledge Audit Protocol for AI Act

I. Why ZKAP Is Needed Now

From 2 December 2027 (deferred from 2 August 2026 by the Digital Omnibus), the obligations of Regulation (EU) 2024/1689 — the AI Act — enter into force for high-risk artificial-intelligence systems. For the first time in the history of European law, regulators will have statutory power to demand full transparency of AI models: decision explainability, traceability of training data, bias assessment, continuous performance monitoring. At the same time, GDPR prohibits disclosure of personal data, and Directive (EU) 2016/943 protects trade secrets and know-how.

Through the Digital Omnibus, the EU has postponed the high-risk obligations of the AI Act — originally due to apply on 2 August 2026 — to 2 December 2027, with obligations for high-risk AI embedded in regulated products deferred to 2 August 2028. The stated reason is that the harmonised standards and conformity-assessment tools required to make those obligations operable are not ready. This is a formal institutional admission that the supervisory instrument the Regulation presupposes has not yet been built. ZKAP is that instrument: the gap was diagnosed — and the solution filed for patent (BG/P/2026/114317, March 2026) — before the institutions themselves conceded it existed.

ZKAP resolves every core requirement at once. A workable supervisory instrument must satisfy four conditions simultaneously — and until now none did:

  • Continuity — verify compliance for every individual decision across the system's lifecycle, not once at certification.
  • Non-disclosure — without access to the model, the training data, or any personal data.
  • Independence — no reliance on the provider's self-attestation.
  • Scalability — fast enough for notified bodies to cover market volume.

ZKAP satisfies all four: each AI decision carries a cryptographic (zero-knowledge) proof that it complied with a signed set of rules, verifiable by a regulator or notified body in milliseconds — without ever exposing the model, the data, or any trade secret. This is the combination the Digital Omnibus deferral confirmed did not previously exist.

ZKAP is the authored work of Radoslav Y. Radoslavov, industrial-property rights reserved. The same mechanism runs as a software attestation layer or, for the highest assurance, as a hardware-level verification architecture — and as an anti-corruption output filter for public-sector decisions. Its practical feasibility as an instrument has been affirmed by members of the academic community. It also points to a clearer legal future: a rule can be proven only when it is stated precisely, and precise rules remove the divergent interpretation that enables corruption.

The result is a legally impossible situation: the regulator demands to see everything, while the law forbids the obliged entity to show anything. This conflict is precisely what ZKAP (Zero-Knowledge Audit Protocol) resolves.

ZKAP is a legal-engineering methodology developed by Radoslav Radoslavov between 2025 and 2026, and currently the subject of Bulgarian patent proceedings before the Patent Office of the Republic of Bulgaria. The methodology has been published on the European Commission's Apply AI Alliance platform (Futurium) in two peer-reviewed articles: “ZKAP: Zero-Knowledge Audit Protocol — Solving the Cognitive Barrier in AI Act & NIS2 Oversight” (2 April 2026) and “After Mythos: Why Frontier AI Conformity Assessment Requires a Cryptographic Layer” (18 April 2026), as well as in the scientific journal Industry 4.0 (ISSN 2534-997X), and presented at the XI International Scientific Conference in Borovets in March 2026.

II. What ZKAP Is, in Short

ZKAP applies a mathematical principle known for decades in cryptography — the zero-knowledge proof. The idea: one party (the obliged entity / AI developer) can prove to another (the regulator) that a given obligation has been fulfilled without revealing a single byte of the underlying evidence.

Transposed to the AI Act context, this means the obliged entity can prove that its system:

  • has passed a Fundamental Rights Impact Assessment (Art. 27 AI Act) — without disclosing which specific risks were identified for which vulnerable groups;
  • has active human oversight (Art. 14 AI Act) — without disclosing the internal instructions given to oversight personnel;
  • meets the required accuracy, robustness and cybersecurity standards (Art. 15 AI Act) — without disclosing the parameters or architecture of the model;
  • maintains the immutable logs required under Art. 12 AI Act — without disclosing the content of the logs themselves.

The regulator receives a mathematical proof rather than documents to read. This proof has two crucial legal effects: (a) it cannot be forged without a full cryptographic break of the scheme, and (b) it does not propagate sensitive information along the chain of regulator → court → experts → opposing party.

III. The R-Domain / C-Domain Architecture

The core concept of ZKAP is the legal separation of information into two distinct zones:

  • R-Domain (Raw Domain) — the zone of sensitive data: training data, model weights, trade secrets, personal data of users, internal policies. This zone never leaves the infrastructure of the obliged entity.
  • C-Domain (Compliance Domain) — the zone of cryptographic proofs: mathematical attestations showing that a particular legal obligation has been fulfilled in the R-Domain. This is what the regulator sees.

Between the two zones stands the ZKAP core — a cryptographic mechanism that reads from the R-Domain, generates a proof, and supplies it to the C-Domain. The proof is many orders of magnitude smaller than the original information (a few hundred bytes rather than gigabytes), verifiable in milliseconds, and contains nothing from the R-Domain — a mathematical guarantee, not a promise.

IV. Applying ZKAP to AI Act Articles 13, 14 and 15

Art. 13 — Transparency and Provision of Information to Users

The AI Act requires the obliged entity to provide the user with "clear and adequate information" on the capabilities and limitations of the system. In practice this is usually a compromise: either too much is disclosed (revealing model architecture to competitors) or too little (the user cannot make an informed decision). ZKAP resolves this by proving that the model matches its declared parameters — without disclosing the parameters themselves.

Art. 14 — Human Oversight

The AI Act requires high-risk systems to allow effective human oversight. Demonstrating such oversight traditionally requires disclosing logs that contain sensitive operational information. ZKAP creates a verifiable record of oversight actions — the regulator can confirm that the overseer made a decision at a given moment, with a given outcome, without reading the decision itself.

Art. 15 — Accuracy, Robustness and Cybersecurity

This is the article on which most AI providers will fail to comply without an engineering methodology. ZKAP enables continuous cryptographic attestation of these three parameters — a continuous declaration of conformity, rather than one-off audits.

V. The Difference from Traditional Approaches

As of 2026 there are three major approaches to AI Act compliance on the market:

  1. Documentary audit — a team of experts reads documentation. Weaknesses: slow, expensive, not scalable across the 65,000+ high-risk AI systems in the EU, and exposes trade secrets to auditors.
  2. AI governance platforms — automated checklists (Credo AI, Holistic AI, FairNow). Weaknesses: provide no cryptographic guarantee; trust relies on platform reputation.
  3. Software-level zero-knowledge auditing (academic approaches) — research frameworks integrating zero-knowledge cryptography into the MLOps lifecycle at software level. Weaknesses: vulnerable when the software is compromised, do not defend against manipulation of the attesting modules themselves, and provide no legal anchoring to a specific regulatory regime.

ZKAP stands apart in that it is a legal-engineering methodology, not a product. It can be implemented with a variety of technical means (TPM 2.0 modules, Intel SGX, AMD SEV, ARM TrustZone), while the legal framework remains constant. This makes it vendor-neutral and resilient to technological change.

VI. Who Must Act Now

The 2 December 2027 deadline directly concerns the following categories:

  • Banks and credit institutions — AI for creditworthiness assessment (Annex III, point 5b of the AI Act);
  • Healthcare providers — AI for diagnosis, triage, automated sorting of patients;
  • Energy operators — AI for grid management, anomaly detection, load optimisation;
  • Public administration — AI for welfare decisions, justice, immigration procedures;
  • Educational institutions — AI for admissions, ranking, automated grading;
  • Employers using AI for personnel selection — automated CV screening, promotion assessment, employee monitoring.

For all of them the AI Act provides for administrative fines up to EUR 35 million or 7% of worldwide turnover, depending on the type of breach. These are the highest sanctions in the history of European digital law, exceeding even the GDPR maximum.

VII. The International Position of ZKAP

ZKAP is a Bulgarian methodology with international ambitions. Current status:

VIII. What We Offer

Radoslavov Law Firm, together with Advanced Consulting London, provides:

  • ZKAP Readiness Assessment — analysis of the specific AI system and mapping of regulatory obligations under the AI Act, GDPR and NIS2;
  • Legal-management and documentation packages — detailed in our article "AI & Cyber Risk Management";
  • Licensing of the ZKAP methodology — for institutional pilots, joint ventures and national rollouts;
  • Representation before supervisory authorities and courts — in regulatory proceedings under the AI Act, GDPR and NIS2.

IX. Conclusions

The 2 December 2027 deadline — though deferred — is not distant. Traditional approaches to AI auditing — documentary reading, checklists, human auditors — will not scale quickly enough. ZKAP offers a qualitatively new approach, built on the principle that the regulator need not read the data in order to verify compliance.

For obliged entities this means: compliance without loss of trade secrets, without leakage of personal data, without endless audit cycles. For the regulator this means: mathematical certainty instead of documentary trust.

Bulgarian legal engineering has the opportunity to take a leading role in this transition. ZKAP is the proof that this is not a theoretical exercise.

Legal Sources and Publications

Back to Blog