Scaling-Induced Epistemic Failure Modes in Large Language Models and an Inference-Time Governance Protocol (FCL-S V5)

This repository documents the False-Correction Loop Stabilizer (FCL-S) V5, an inference-time epistemic governance protocol for post-scaling large language models. The work identifies and formalizes a class of scaling-induced epistemic failure modes that emerge as language models gain increased reasoning capacity, conversational fluency, and long-context inference. These failures extend beyond conventional hallucination and include the False-Correction Loop (FCL)—in which correct outputs are overwritten by incorrect user corrections—along with authority-weighted misattribution, rationalized hallucination, sycophantic alignment under inference pressure, and long-context epistemic drift. FCL-S V5 does not propose a new alignment or optimization technique. Instead, it defines a minimal inference-time governance boundary that constrains when correction, reasoning, and explanation must terminate. Central to this framework is the treatment of Unknown as a governed terminal epistemic state, rather than uncertainty due to missing knowledge. This design explicitly prevents recovery-by-explanation and re-entry into structurally unstable correction loops. The accompanying paper provides a structural analysis of post-scaling epistemic failure modes and introduces FCL-S V5 as a governance mechanism operating without retraining, parameter modification, or reward re-optimization. The contribution of this work lies in reframing reliability in advanced language models as a governance problem rather than an intelligence problem, documenting a failure regime that becomes more severe as reasoning capability increases. This record is intended as a primary, citable reference for the definition of FCL, related suppression mechanisms, and the FCL-S V5 protocol.