TU/TUE/CIGL: Towards Epistemic Integrity, Accountability, and Failure & Risk Visibility in AI
Year: 2025
Type: article
Abstract: Dominant AI paradigms, particularly connectionist models (DL, LLMs), often exhibit significant epistemic deficits due to inherent opacity and lack of intrinsic epistemic scaffolding, hindering verifiable understanding and control. This paper introduces the Transparence Unit (TU) Ecosystem (TUE), supported by the Cognitive Integrity Glue Layer (CIGL) infrastructure, as a conceptual architecture promoting intrinsic epistemic governance and risk visibility in AI. The Transparence Unit (TU), evolving from EpistemicState, is proposed as a universal, auditable "cognitive packet" designed for embedded traceability, explicit dependency tracking (including scaffolding references), and mandatory justification relative to versioned external resources. TUs function within the TUE, managed by the CIGL runtime and governed by specialized agents (FCJF, G-Layers, AIRE, ELIM, TRACE). The framework addresses epistemic weaknesses via mandatory justification (FCJF), staged maturity (G-Layers G7-G0), governed lifecycles, and integration pathways for opaque models requiring external validation via "epistemic scaffolding". The vision aims for systemic governance where epistemic risk profiles and justification pathways become operational properties. A phased roadmap (Proto -> Ω -> ∞) outlines development from passive logging towards operational and pinnacle governance. This conceptual architecture requires further research and validation but aims to provide a foundation for AI systems demanding verifiable integrity, accountability, and justifiable reasoning.
Author: Ivliev, Ihor
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6893065733
FWCI:
Cites:
Cited by:
Related to: