Research &
Publications

Abstract

We introduce Pool-Gated Retrieval (PGR), an architecture that reconceives retrieval-augmented generation as a three-stage epistemic process — Horizon, Warrant, and Commitment. By treating absence as a first-class fact, PGR registers confirmed absences as structural facts and propagates them to the answer, addressing absence hallucination by structural design, and every answer carries an auditable provenance record. Demonstrated on a single synthetic partial-coverage scenario against a naive RAG baseline.

Key Contributions

• •Three-stage Horizon-Warrant-Commitment retrieval architecture
• •Structural registration of Knowledge Gaps to prevent absence hallucination
• •Auditable provenance record for every committed fact
• •Taxonomy of eight RAG failure modes, with an open Python reference implementation

Abstract

Large language models exhibit intelligence without genuine epistemic understanding, revealing a missing epistemic architecture. This paper introduces the Structured Cognitive Loop (SCL) as an executable epistemological framework, asking under what conditions cognition emerges rather than what intelligence is. Intelligence is reframed as a performed process — a continuous loop of judgment, memory, control, action, and regulation.

Key Contributions

• •"Executable epistemology": philosophy operationalized as structural experiment
• •Five structural conditions for understanding: evidential grounding, memorial persistence, normative control, environmental coupling, recursive self-reference
• •Intelligence redefined as the capacity to reconstruct one's own epistemic state

Abstract

We introduce the Structured Cognitive Loop (SCL), an agent architecture that separates cognition, control, action, and memory. The language model only proposes actions, a lightweight controller validates every proposal before execution, and memory is kept externally as a structured store of verified facts. Across 360 episodes in three controlled scenarios, SCL reached 86.3% task success versus 70.5–76.8% for prompt-based baselines using the same model and tools.

Key Contributions

• •Modular loop: Cognition → Control → Action → Memory, with Retrieval at turn onset
• •Controller-mediated execution with an auditable trace of proposals and approvals
• •Controlled evaluation against ReAct and LangChain agents, with ablations of memory and control
• •Extended revision: human-in-the-loop, Pool-Gated Retrieval, and the Horizon-Warrant-Commitment framework

Abstract

A position paper reporting that a practical LLM agent architecture, Agentic Flow, structurally echoes four theories of mind — Kahneman's dual-system theory, Friston's predictive processing, Minsky's society of mind, and Clark's extended mind. The convergence emerged from implementation demands rather than deliberate synthesis, and is described as a partial reflection, not a unification of the theories.

Key Contributions

• •Structural mapping of one five-module architecture onto four theories of mind
• •PEACE: a descriptive meta-architecture of recurring design patterns, later formalized as SCL
• •Controlled comparison with LLM-only agents: 95.8% vs 62.3% task success

Abstract

A theoretical follow-up asking whether hallucination could be eliminated if lossless abstraction were possible. Drawing on Kolmogorov complexity, Chaitin's incompleteness theorem, and Shannon's entropy limit, it argues that any sufficiently intelligent system must rely on lossy compression, so hallucination is a structural feature rather than a flaw to engineer away.

Key Contributions

• •Extends the compression–generalization–hallucination triad with a lossless-abstraction counter-hypothesis
• •Information-theoretic argument from Kolmogorov, Chaitin, and Shannon
• •Proposes hallucination-aware architectures and epistemically humble interfaces

Abstract

We argue that hallucination is not an incidental flaw but an unavoidable structural property of generative intelligence: generalization rests on lossy compression, and reconstructive inference fills the gaps. Combining theoretical modeling with a synthesis of recent empirical work and parallels to human reconstructive memory, the paper concludes that attempts to eliminate hallucination may dismantle the mechanisms behind generalization.

Key Contributions

• •Three hypotheses: inseparability, memorization–generalization trade-off, creativity–error duality
• •Logical-theoretical argument grounded in information theory and cognitive science
• •Implications for AI safety, sovereign AI policy, and high-stakes domains such as law, medicine, and national security