The deployed bundle is produced offline from the notebook and artifact scripts. Runtime scoring uses the full-history ensemble when a GHArchive cache row exists and the cold-start ensemble when it does not; missing artifacts are configuration errors.

Run the offline notebook or `npm run ml:bootstrap:foundation -- --gharchive-source <path>` to export staged artifacts for this page.

Once a completed run produces evaluation bins, the reliability curve will render here from the cached artifact.

Quality

A held-out summary score that combines AUROC skill with Brier skill. It is useful for comparison, not as a standalone proof.

AUROC

How well the model ranks riskier dependencies ahead of less risky ones across thresholds. Higher is better.

Brier score

A calibration-sensitive probability error metric. Lower is better, which is why it matters for thesis-style reliability claims.

ECE

Expected calibration error compares predicted probability bins with observed outcomes. Lower means the probabilities are better calibrated.

Inactive 12m rate

The share of positive labels in the evaluation slice. It tells you how much true inactivity pressure is present in the held-out set.

F1

The balance between precision and recall once a classification threshold is chosen.

Precision

Of the dependencies flagged as risky, how many truly belong in that slice.

Recall

Of the truly risky dependencies, how many the model successfully catches.

Calibration

Whether predicted probabilities match real observed rates. Good calibration makes a score easier to trust in triage.