ContextFidelity-Bench: A Three-Paradigm Framework for Evaluating Multi-Turn Hallucination Behavior in Language Models. TBench 2026, 6. https://doi.org/10.66834/0azb3w20.