Kim, Myung Ho (2026) Epistemic Architecture of Accountable Artificial Agents: Justified Acceptance Paths and the Horizon-Warrant-Commitment Framework. [Preprint]
![[img]](https://philsci-archive.pitt.edu/style/images/fileicons/text.png) |
Text
Epistemic Architecture of Accountable Artificial Agents.pdf Download (900kB) |
Abstract
Contemporary AI agents exhibit high productivity but lack epistemic accountability: they can produce correct outputs without there being any principled answer to the question of why a given judgment was permitted. This gap is not merely a philosophical concern — it is a social one. When AI agents act on behalf of institutions, societies require not just correct outputs but legitimate decisions: evidence of what was considered, under what conditions it was admitted, and where responsibility lies when things go wrong. This paper argues that closing this gap requires a reconception of agent architecture from the ground up. We analyze the Structured Cognitive Loop (SCL) framework as an implementation of three structurally distinct justified acceptance paths: Memory Fact, grounded in causal tool observation; Contextual Human-in-the-Loop (HITL), grounded in authoritative human judgment; and Pool-Gated Retrieval-Augmented Generation (RAG), grounded in threshold-governed evidential selection. All three paths are unified by the Grounding Principle, which fixes the epistemic basis of any reasoning episode at turn onset and prevents circular epistemic drift. The paper's central theoretical contribution is the tripartite framework of Epistemic Horizon, Warrant, and Commitment. This framework was not imported from epistemological theory and applied to SCL; it emerged from philosophical analysis of SCL's actual operation — specifically, from the observation that Pool-Gated RAG's two-stage architecture enacts a three-stage epistemic sequence that, upon inspection, also governs tool execution and HITL judgment. H-W-C is thus a structure that SCL's own operation made legible, not a schema imposed upon it. The paper further argues that this structure constitutes a general formal condition: any epistemic input can be incorporated into an accountable artificial agent if and only if its Horizon, Warrant, and Commitment conditions can be structurally specified. This positions SCL not as a closed architecture but as an epistemically open system — one that functions less like a particular agent design and more like an operating system for epistemic cognition, whose interface is a formal requirement rather than a fixed list of supported input types. The central thesis is that intelligence appropriate for accountable artificial agents is normative rather than descriptive: it is not a matter of what the agent can produce but of what the agent is entitled to produce, and under what structural conditions that entitlement is conferred.
| Export/Citation: | EndNote | BibTeX | Dublin Core | ASCII/Text Citation (Chicago) | HTML Citation | OpenURL |
| Social Networking: |
Monthly Views for the past 3 years
Monthly Downloads for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](http://philsci-archive.pitt.edu/style/images/feed-icon-32x32.png){kind=icon}