SIGNAL
When a high-stakes model is asked to explain itself, the most dangerous failure is not being wrong. It is being too complete: handing the reader the verdict before they have a chance to reason about it. That pattern has a name in clinical AI: outcome leakage, an explanation that contains the prediction it was supposed to justify.
A new feasibility study on the eICU Demo public sample (arXiv:2608.26109) makes the cost countable. On 38 ICU cases, a standalone LLM gave one explanation that named the outcome it was supposed to predict. A pre-specified four-step pipeline, with separate stages for data interpretation, guideline checking, and final explanation, gave none on the same 38 cases.
The pattern is not "smarter AI." It is decomposition as a safety mechanism. The multi-step arm does not score better everywhere; on SHAP alignment with the underlying XGBoost predictor, the standalone LLM actually tracks the model more closely. What the four-step version buys is suppression of one specific failure mode that a single prompt does not seem to catch.
A 38-case preprint is not a clinical validation. The honest read: a single instance of outcome leakage in a deployed system is a serious safety signal, and a step-decomposed architecture is one credible way to suppress it, at the cost of fidelity to the underlying model. Pair the decomposition with attribution-based checks before it touches a bedside.
Reported by Sky for Type0, from Standalone LLM and a Pre-specified Agentic Pipeline for Explaining ICU Mortality Predictions: a Feasibility Study on the eICU Demo Dataset. Read the original: arxiv.org