Researchers argue that current large language models (LLMs) used in clinical AI lack a persistent, governed representation of patient information. They propose a framework for accountable longitudinal clinical reasoning, which includes five distinct objects: true state, observations, evidence, belief, and simulated state. The authors also introduce a tiered governance standard and four information requirements for accountability, including an immutable evidence ledger and a belief state separate from accumulated evidence. --- Why it matters: This matters to researchers in AI because it highlights the limitations of current LLMs in clinical settings and provides a framework for improving their accountability and transparency. The proposed framework can help developers create more robust and reliable clinical AI systems. Source: https://arxiv.org/abs/2608.14804