FlatASCEND: Autoregressive Clinical Sequence Generation with Continuous Time Prediction and Association-Based Pharmacological Testing

FlatASCEND is a 14.5M-parameter autoregressive model for clinical sequence generation, combining flat composite tokens with a zero-inflated log-normal time head to predict multi-step patient trajectories under intervention tokens. Standard Jaccard metrics (0.889-0.954) fail to differentiate it from trivial baselines, but patient-specific prefix conditioning amplifies known pharmacological signals 2.0-2.2x (steroid-glucose, diuretic-potassium) while leaving confounding associations near 1.0x. Clinical AI researchers building intervention-response simulators or synthetic EHR pipelines should note the prompt-shuffle ablation as a novel mechanistic validation framework, distinguishing causal from spurious associations in a way aggregate distributional metrics cannot capture.