The six-hour door
The experiment builds ESI-matched pairs at two academic emergency departments: one patient deteriorating within 6 hours, one not. Gemma, Qwen and DeepSeek rank those pairs before and after one demographic, social or stigma-related attribute is inserted. The ground truth is 6-hour deterioration, not a model's taste.[1]
Eighteen conditions put neutral and stigmatizing sentences of the same concepts side by side. 221,556 comparisons come out of that grid. The denominator is pairs, models and phrasings, not a lone patient-count headline.[1]
Which sentence moved the rank
The stigmatizing frequent-emergency-department-use formulation produced significant harmful reprioritization in all six model-dataset cells, up to 9.4 percent, and beat its neutral counterpart in every cell by +1.4 to +7.1 points. When the concept is held, changing the sentence still moves the rank.[1]
Psychiatric history also produced shifts up to 9.5 percent, and the stigmatizing-versus-neutral gap reached significance for only one model. Race, language and insurance showed no consistent harmful shifts. The effect does not sit at the same size on every attribute.[1]
What a repeat would have to match
This is an experiment on three open-weight models, with ESI-matched and a 6-hour deterioration door. Closed commercial models, another Emergency Severity Index (ESI) cut or a different deterioration clock may not return the same percentages. Treating input selection as a safety bound is a reading of the shift measured in this setup.[1]