Accuracy Without Stability Is Not Intelligence: Behavioral Evaluation as the Missing Dimension in LLM Benchmarks

This paper demonstrates that output correctness is a necessary but insufficient indicator of large language model reliability. Through 2,040 instrumented observations across two production models and a 27-model behavioral benchmark, we identify failure modes invisible to accuracy-based evaluation: models producing correct outputs while operating in unstable internal states, reasoning that changes under minimal perturbation, and self-contradiction under verification. We propose three observable behavioral tests — Consistency, Continuity, and Self-Agreement — requiring no access to model internals. Companion paper to the Cognitive Entropy Collapse Battery (CECB) Kaggle competition submission. Published figures embedded in HTML.