KC-Bench: no model handles factual correction, identity consistency and temporal conflict at once.

238 tasks screened from 1,000+ candidates, measuring how agents reconcile user instructions, parametric knowledge and live environment observations. Across nine models including DeepSeek-V4-Flash, GLM-5.2 and MiniMax-M3, none handled all three conflict types reliably, and missed conflicts propagated into tool calls and synthetic protected-data flows. Usable as a model-selection diagnostic when your tools return data that can contradict the prompt. arXiv 2609.03588