LAB / 02
Model explorer
The seven places a scoring model can quietly go wrong, laid out in order: where the data was split, what was fitted on what, how training went, how the model was chosen, what the locked test says, whether the population moved, and what a threshold costs.
Demo predictions come from templar-fraud-lab; the drift panel from model-monitoring-lab. All synthetic.
1 · Partition timeline
Chronological split
Loading trace
2 · Preprocessing fit scope
What was fitted where
Loading log
3 · Training progress
From the training log
Loading log
4 · Validation selection
How the model was chosen
5 · Locked test
Touched once, at the selected operating point
Loading metrics
6 · Drift
Population stability by window
Loading monitoring
7 · Demo threshold
Recompute queue totals and cost from the demo predictions
Loading demo predictions