Ario：面向 AI 系统认知完整性审计的初步设计与探索性评估

Published October 9, 2026 | Version v7

Description

This preliminary research report introduces Ario, a proposed framework for auditing epistemic integrity, memory, revision, and identity-related claims in AI systems. It presents the project's foundational principles, including the distinction between memory and truth, the separation of evidence from inference, explicit representation of uncertainty, and accountability for revisions.

The report includes an exploratory evaluation of conversational transcripts involving NoTrack, focusing on evidence-grounded retrieval, source independence, cross-session continuity claims, integrity and immutability claims, repeatability, and self-correction. The observed interactions reveal both useful epistemic behaviors and significant failure modes, including unsupported assumptions, fabricated reconstructions of prior assessments, and corrections made only after explicit challenge.

The evaluation is limited to user-provided transcripts. It does not independently verify NoTrack's internal architecture, persistent memory, technical audit mechanisms, or implementation of Ario. The findings are therefore preliminary observations rather than proof of system-wide capabilities.

The report establishes an initial research direction for developing and testing auditable AI systems in which claims, evidence, provenance, uncertainty, and revisions can be examined without granting any actor automatic epistemic authority. It is intended as an early design and exploratory evaluation document that can be refined through reproducible experiments, explicit test specifications, and independently inspectable technical evidence.

Project repository: https://github.com/SepehrGhanbari/Ario

Files

Preliminary Design and Exploratory Evaluation.pdf

Files (187.0 kB)

Additional details