What was known?
Trace three model features to their source records, revisions and publication or ingestion times. A synthetic cohort of 480 applications includes late historical corrections.
Inspect an application ↗Credit scores.
With evidence.
From the records available at a decision to mature credit outcomes. Inspect the inputs, the evaluation and the assumptions behind each result.
Does it hold up?
Separate training, calibration and held-out periods. Compare raw and calibrated probabilities, vintage results and a fixed review-capacity scenario.
Open the original risk report ↗Has the score drifted?
Inspect the score distribution before new credit outcomes mature. Follow the reference bins and application-level evidence behind the diagnostic.
Explore score monitoring ↗Synthetic research examples. The source-lineage example and the original risk report use separate cohorts. The original report retains the measured worsening of Brier score after independent isotonic calibration. These examples do not describe real borrowers or a deployed credit policy.