What this does: loads the same
cohort_graph.json from Steps 2–4 and layers three independent filter axes over it,
without deriving any new fields or re-ingesting data: a population filter (T1D / T2D / all), a disease-appropriate
intervention-timing filter (early vs. delayed insulin initiation for T2D; pump-failure evidence vs. none for T1D — the same
approximation flagged in Step 4), and a comparator matrix of outcome (complication-free / complicated), facility tier, and
resource status. Each filter only constrains the population it applies to and passes other patients through unfiltered, so mixed T1D+T2D
views stay meaningful. Results can be grouped by patient, by facility tier, by resource status, or by time-to-first-complication bucket.
No file loaded yet — use the cohort_graph.json exported from Step 2 (same file works for Steps 3–5).