Every plant carries a probability that it suffers a disruption in the next 365 days — a failed FDA inspection or a shipment refused at the border for manufacturing quality. It is learned from the federal record, not assigned.
Trained on
every drug-establishment inspection and how it ended
Oct 1, 2008 → Aug 4, 2026 · 18,887 establishments · 90 countries
- NAI (no action)17,957
- VAI (voluntary action)18,716
- OAI (official action)3,002
shipments turned away at the US border
Jan 2, 2014 → Aug 26, 2026 · 7,788 establishments
- manufacturing quality (charge 27, 3280)7,701
- paperwork (charge 118, 472)5,980
who is registered to manufacture, and where
The fit
- Plants7,330
- That failed within a year172
- Base rate2.3%
- CutoffAug 1, 2025
- Features9
Label: a failed inspection (OAI) or a manufacturing import refusal in the 365 days after the cutoff.
Why that cutoff: The most recent date with a full year of outcomes behind it. Anything later and the label is only partly observed, which reads as a lower risk than is real.
Does the percentage mean anything?
Rank accuracy cannot answer that — a model can order plants perfectly and still be wrong about magnitude. So each plant is scored by a model that never saw it, and the predictions are compared against what actually happened.
| Predicted | Observed | Plants | |
|---|
| 0.46% | 0.57% | 3,688 | ✓ |
| 1.6% | 2.2% | 1,456 | ✓ |
| 2.5% | 1.9% | 1,264 | ✓ |
| 6.8% | 6.0% | 604 | ✓ |
| 13.5% | 13.2% | 257 | ✓ |
| 35.4% | 41.0% | 61 | ✓ |
Every band lands inside the 95% interval of what was observed. Nothing is shown above 46.0% — the highest rate ever actually observed at any score.
