External Verification Architecture (EVA)
The Structural Foundation for AI Safety
External Verification Architecture (EVA) is a technical framework for continuously, structurally, and verifiably monitoring the reasoning processes of AI systems through mechanisms independent of the model’s internal parameters.
EVA enables AI safety to be proved, not merely claimed.
Three Structural Conditions
1. Verifiability
Each reasoning step is output externally in machine-readable format (e.g., JSON‑LD / RDF).
2. Transparency
A monitoring mechanism operates in parallel with inference and cannot be influenced by the model.
3. Physical Immutability
Verification logic is implemented on a substrate that cannot be altered by software (FPGA / ASIC / TPM).
These three conditions form the minimum structural requirements for genuine external verification.
Purpose of EVA
EVA provides a structural foundation for:
Continuous verification of AI reasoning
Independent third‑party oversight
Post‑incident analysis and corrective action
Compliance with international AI safety regulations
EVA shifts AI safety assurance from assertion to proof.
Reference Materials
Book
External Verification Architecture (EVA): The Missing Foundation for AI Safety
ISBN: 9798196129889
Author: Satoru Hara
Publisher: Natural Structure Works (NSW)
Published: May 2026
DOI (Zenodo / SSRN)
GitHub: https://github.com/satoru-hara/03_NSW (github.com in Bing)