Public demoAggregate findings come from the real MIMIC-IV cohort run. Patient-level fields are redacted per the PhysioNet DUA. Inference was performed locally during the original run. Candidate table is capped at 50 rows per state for this public demo.