Illustrative reference pipeline. The real model pipeline runs off-platform (arm64) via make -C ml/bthal all — see the "From cohort to prediction" steps and /provenance for the real lineage + run record.
reference DAG
M0 … M16a
illustrative modulesmodules shown
7 / 17
quality gates
10 passed · 0 failed · 14 pending
fitted model
XGBoost · per endpoint
ml/bthal
Pipeline DAG · 20 M-prefixed modules · local-mac
run_0061 · live
completerunningqueuedblockedhelper (not counted)
Runtime components · served alongside the build pipeline
excluded from build-DAG count and quality-gate KPI
M13rruntime
Predict service
fastapi mock · batch
SLO · p95 latency < 200 ms (batch=32, 1k requests)
Fitted model
one calibrated XGBoost per endpoint
arm64 · synthetic-labeled cohort · no PHI
Contract
validation · figures · predictions
committed under public/data/bthal/
Reproduce (off-platform)
make -C ml/bthal all
cohort → train → evaluate → serve
From cohort to prediction
- 01generate_cohort → synthetic-labeled cohort (base rates → published anchors)
- 02train → one XGBoost classifier per endpoint + XGBoost regressors
- 03calibrate → isotonic / Platt by ECE
- 04evaluate → AUROC/AUPRC/Brier + calibration + DCA + TreeSHAP
- 05export → public/data/<id>/* contract (validation, figures, predictions)
- 06serve → FastAPI /predict (heuristic fallback if the server is down)