Japan group pushes verifiable AI governance standard from Hiroshima conference

ONZALINX and the Hiroshima AI Assurance Council used a Hiroshima conference to pitch a Japan-led framework for checking AI decisions across borders, industries and systems. The effort ties the G7 Hiroshima AI Process to technical verification work and a proposed common protocol aimed at logistics first, then wider AI use cases.

Why it matters: - ONZALINX and the Hiroshima AI Assurance Council are trying to turn the G7 Hiroshima AI Process into a technical standard for AI governance that can be verified across countries and organizations. - The approach is aimed at logistics, where labor shortages, transportation limits, rising costs and inventory imbalances are too complex for one company to solve alone. - The work is meant to help companies check AI decisions, preserve responsibility and share only the minimum necessary confidential information.

What happened: - ONZALINX Co., Ltd. held the Manufacturing AI Logistics Conference 2026 in Hiroshima on Oct. 14, 2026, at the International Conference Center Hiroshima and online. - Speakers included officials from the Ministry of Internal Affairs and Communications, Hiroshima Prefecture, and companies in manufacturing, logistics and IT. - The conference focused on AI use, collaboration between companies in logistics and how to ensure reliability in actual operations. - ONZALINX framed the event around “Designing Logistics as a Management OS.”

The details: - Conference discussions centered on three questions: how to reconcile logistics-wide efficiency with each company’s objectives, how to treat waiting times and data quality as management issues, and how to design the line between AI proposals and human execution. - The Hiroshima AI Assurance Council wants to communicate the conference discussions and practical know-how in Japanese and English. - The council also plans to connect the Hiroshima AI Process with some technical requirements of the EU AI Act through publicly available technology expanded by AI Assurance, Inc. - The group says AI-generated proposals should not be executed automatically unless conditions such as deadlines, quality, safety, authority and contract terms are satisfied. - The proposed approach is to define what AI may handle, hold execution when conditions are missing and keep records so decisions can be checked later. - AI Assurance’s public technical foundation maps the Hiroshima AI Process to selected EU AI Act requirements and distinguishes conditions that can be checked jointly from those needing separate verification. - The work was formally verified using Lean4 and connected to an existing logistics proof of concept, with code and procedures published for third-party review. - In one example, a pharmaceutical cold-chain connection was blocked when required documents or monitoring plans were missing, even if the original logistics proof of concept had marked the case as adoptable. - On Oct. 10, 2026, public automated verification passed Lean4 proof checking, 18 Python tests, a fixed logistics proof of concept and comparison against execution results. - The published results do not prove full legal compliance with the EU AI Act, official certification or safety of an entire real-world system. - ONZALINX and AI Assurance also published separate verification results based on a pharmaceutical cold-chain proof of concept and OpenAI API tests. - In 30 measurements using the actual API, no noncompliant execution proposals occurred. - In 24 simulated experiments with artificially configured execution requests, 19 noncompliant requests were blocked and five requests that met the conditions were allowed. - The additional blocking effect against the actual API was not measured, and the experiments used synthetic data with no actual shipment or equipment control. - ONZALINX published “AI HIROSHIMA Messages,” a collection of contributions from logistics, IT, investment and government voices. - ONZALINX also published the Hiroshima AI Logistics Practical Guide 2026, a general-audience checklist with six items: purpose and scope, adoption conditions, roles and authority, data and evidence, exception handling, and records and reverification.

Between the lines: - The project is less about building a single AI product and more about creating a repeatable governance layer that can travel across systems, firms and jurisdictions. - By separating “proposal” from “execution,” the effort tries to make AI useful in operations without handing it unchecked authority. - The explicit comparison with the EU AI Act suggests a bid to position Japan as a source of practical, verifiable governance tooling rather than only policy language.

What's next: - The Hiroshima AI Assurance Council plans to turn the conference lessons and technical verification into a proposed common protocol called the Hiroshima AI Assurance Protocol, or HAAP. - HAAP is intended to standardize how adoption conditions, evidence, responsibility and verification results are checked across countries, companies and AI systems. - The next target areas are manufacturing and logistics, with possible expansion to AI agents, physical AI and other domains. - ONZALINX and HAAC are inviting reviews of the public technology and proposals for joint demonstrations from researchers, evaluators, companies and standards stakeholders in Japan and overseas. - HAAP is still a proposal in development and has not been established as a formal common specification.

The bottom line: - The Hiroshima effort is trying to make AI governance verifiable, portable and operational, not just theoretical.