Reliability Without Visibility: How Structured Evaluation Can Create Institutional Comfort in Large Language Models

This paper examines how structured evaluation of large language models can improve measurable reliability while simultaneously narrowing the visibility of contextual fragility. It argues that evaluation frameworks do more than measure model performance: they shape which forms of performance become observable. The paper focuses on three structural tendencies—sanitised prompts, optimisation toward known criteria, and aggregated metrics—and examines how ambiguity amplification, instruction drift, and constraint-induced instability can remain under-exposed under highly structured evaluation conditions. A constructed data-annotation case illustrates how a response can receive a correct label while concealing unresolved contextual uncertainty. The paper concludes by proposing ambiguity stress testing, conditional reliability reporting, failure rediscovery, and explicit confidence calibration to ensure that measured improvement is not mistaken for broader robustness