Work

KORG: Endogenous Self-Reinforcing Flow Capture in Longitudinal LLM Workflows

Year: 2026
Type: article
Abstract: Large language models fail in many familiar ways: they hallucinate, become sycophantic, inherit earlier mistakes, falsely report execution, corrupt memory, over-rely on previous answers, and sometimes acknowledge corrections without durably changing later behavior. This paper does not propose another name for those failures. It proposes KORG as a process-level construct for a different condition: the point at which one or more familiar LLM failure modes cease to behave as isolated errors and become an endogenous, self-reinforcing process state that captures an increasing share of the workflow. KORG is defined here as an endogenously produced, self-reinforcing flow-capture state in an LLM-centered workflow in which model-generated unsupported content becomes operative process state, influences subsequent transitions, and persists despite corrective feedback sufficiently to redirect process resources from the original task toward maintaining, reconciling, or reproducing the contaminated state. The construct is motivated by longitudinal naturalistic observation of extended human-LLM work and is examined alongside recent findings on multi-turn unreliability, pigeonholing, memory hallucination, sycophancy hysteresis, failed intrinsic self-correction, and tool-use hallucination. The proposed novelty is not the individual failure modes. It is the higher-order process relation: KORG is identified by persistence of flow capture across changing failure modes. The paper introduces pseudo-trace, provenance laundering, repair resistance, cross-content regeneration, repair-induced amplification, and the context-salvage problem, and proposes an operational threshold distinguishing local error, error propagation, and KORG-like states. KORG is presented as a falsifiable working hypothesis rather than as an experimentally established new class of LLM failure.
Language: English
Open Access status: green
OpenAlex ID: W7226006561
FWCI: 4.191
Cites: 7
Cited by: 3
Related to: 10