AI Auditing: From "Trust Us" to "Mathematically Verifiable" — GhostDrift Releases Formal Proof of ADIC

Saving "evidence of what AI did" in a way that anyone can reproduce and verify using the same procedure later — this is the third-party verifiable AI governance foundation required by the EU AI Act.

GhostDrift Mathematical Research Institute has released a formal proof, verifiable in Lean 4, for the replay verification theory, which is at the core of "ADIC (Advanced Data Integrity by Ledger of Computation)," a technology that preserves AI judgments as evidence that can be re-executed and verified by third parties. ADIC fundamentally differs from conventional audit logs in that it preserves not only the result of a judgment but also the process leading to that judgment as evidence.

In high-responsibility domains such as logistics, pharmaceuticals, finance, and the public sector, where AI judgments directly lead to accidents, significant losses, or administrative decisions, if it is not possible to explain "why that judgment was passed" after the fact, an "evaporation of responsibility" occurs where no one can take responsibility. ADIC is a foundation that prevents such situations where "everyone was supposed to be right, but a problem occurred," enabling companies and organizations to take definite responsibility for AI.

The structure of "evaporation of responsibility" that ADIC prevents

In Europe, the full application of the EU AI Act, which can impose penalties of up to 7% of the previous year's worldwide annual turnover for serious violations, is approaching on August 2, 2026. The level of demand for AI governance is globally shifting from the conventional stage of operating with strict internal rules to the stage of presenting evidence that can be verified by independent third parties.

The formal proof released this time makes the concept of "replay verification," which is at the core of ADIC, reproducible and verifiable by third parties on Lean 4. This demonstrates that when ADIC's verifier accepts a certificate, the corresponding validity conditions hold for the content indicated by that certificate. Lean 4 is a theorem proving assistant mechanism that allows computers to inspect mathematical proofs. It should be noted that this formal proof targets the soundness of ADIC's replay verification core and does not prove the entire operational system, the conversion from real data to formal models, or completeness. With this release, ADIC has demonstrated a core verification foundation for practical operation as an AI governance technology accompanied by technical evidence that can be re-executed by third parties, rather than just a concept or operational rule.

Development Philosophy "Wasan 2.0": Japanese Thought Supporting International Standards for AI Governance

The design philosophy of ADIC is rooted in the Japanese mathematical culture where "the procedure itself is the basis of trust."

In Wasan (Japanese mathematics) during the Edo period, demonstrating "procedures that anyone can re-execute" by publishing solutions on sangaku (mathematical tablets) was a proof of trust. This was an attitude of defining a "reliably verifiable range" and preserving procedures, rather than discussing infinite theories. ADIC is a modern implementation of this spirit. GhostDrift positions this mechanism, which fixes AI decision procedures as an evidence ledger and allows third parties to mechanically re-execute and verify them, as "Wasan 2.0" and aims to connect it to international standards as a Japanese-born AI governance technology.

Wasan 2.0 — The spirit of publicizing procedures for re-verification of AI judgments

Representative's Comment

Hideaki Maeki, Representative Director, GhostDrift Mathematical Research Institute "It is already a global premise that AI trust can no longer be established solely through ex-post 'explanation' or 'monitoring.' What is important is that after a judgment, a third party can retrace the same steps to confirm what the basis was and where it should have been stopped. What we are implementing with ADIC is the 're-verification responsibility' of AI judgments. The formal proof using Lean 4 released this time demonstrates that ADIC's replay verification core is a technical foundation that can be re-executed by third parties. From governance that is 'just watching' to assurance that can be 'verified later' — GhostDrift implements this transition as a technology originating from Japan."

Future Developments: Implementing AI Assurance Starting from "Hiroshima," the Origin of International Discussions

GhostDrift aims to connect to international standards for "AI Assurance," centered on third-party verifiability and reproducibility, while advancing ADIC implementation in high-responsibility domains such as logistics and pharmaceuticals. AI Assurance, which technically guarantees that AI system judgments are trustworthy and safe, is currently being rapidly developed as a core area of AI governance in the UK, EU, and US.

As a first step towards these international standards, in April 2026, in Hiroshima, the birthplace of the "Hiroshima AI Process" which became the starting point for international discussions on AI governance, GhostDrift concluded a strategic partnership with On-The-Links Co., Ltd., which has a long track record in high-tech manufacturing logistics, and began a PoC for a shipper responsibility certification platform in pharmaceutical logistics cold chains.

As the first milestone based on this partnership, an initial PoC will be conducted over the next two months.