Write a PREreview

Abstract

Large language models are often judged through fluency, helpfulness, and broad task performance. Yet in practically important settings, the central reliability problem is not generation alone but premature release of structurally weak outputs. A model may produce plausible language while still exhibiting unsupported certainty, logical invalidity, ambiguity-insensitive answering, or unsafe behavior in high-burden contexts. This paper proposes a governance-centered reliability architecture that treats model outputs as provisional candidates rather than automatically releasable answers. The framework separates candidate generation from burden-sensitive evaluation, reflective review, and final release decision. At a high level, the architecture incorporates input burden analysis, ethical-logical evaluation, reflective review, and multi-state release control. Instead of relying on a single answer/refusal pathway, the system supports differentiated outcomes such as approval, qualified approval, revision, clarification request, deferment, and rejection. A preliminary mini-pilot using a small instruction-tuned baseline model provides directional evidence in selected burdened cases, including logical invalidity blocking, ambiguity clarification, brittle factual release control, and at least one high-risk defer case. The results do not establish broad statistical superiority, but they do support the central claim that reliability can improve when release is governed rather than delegated to raw first-pass generation. The work is therefore presented as a proof-of-concept architecture with preliminary pilot evidence and a closed-core implementation strategy. Keywords: Large language models; reliability; governed release; admissibility; reflective review; burden- sensitive control; output governance; uncertainty control; hallucination mitigation; AI safety