Beyond Fluency: A Human-Centered Evaluation Rubric for Grounding, Answerability, and Reliability in LLM Outputs

Large language models can produce polished responses whose apparent completeness exceeds their evidential support. Existing evaluation work already measures many relevant properties, including accuracy, factuality, robustness, calibration, helpfulness, safety, and human preference. This paper proposes a complementary human-review rubric organized around three questions that are easy to lose when fluency dominates judgment: Is the response grounded in the task and available evidence? Does it expose uncertainty and correction conditions rather than overstate closure? And does repeated use justify reliance? The framework retains the terms grounding, answerability, and reliability, while mapping them to established work on factuality, calibration, human evaluation, and holistic model assessment. It also develops persuasive over-completion as a failure tag for answers that are more definite, specific, or settled than their support warrants. A three-dimension scoring profile is proposed for human annotation, but reliability is explicitly treated as a repeated-performance property rather than something a single response can prove. The rubric is therefore presented as an instrument requiring validation, not as an established benchmark. A validation plan specifies inter-rater agreement, convergent and discriminant comparisons with existing metrics, stress tests under prompt variation, and user studies of trust calibration. The aim is a practical supplement to technical evaluation: a way to identify when a fluent answer invites more trust than the underlying evidence and evaluation history justify.