Red-team your models before reality does.

Red-teaming was a new phrase to me recently.  I’ve called it a few other things over time: pressure-testing, finding the outliers, “what can go wrong will go wrong”, etc.
Have your programs been red-teamed?  Suicide prevention, FWA, ethics complaints, etc? They all look solid on paper, beautifully-designed SOPs that sit in a manual with colorful arrows and decision trees.
And then real life happens: incomplete information, time constraints, fluctuating acuity, human variability, poor tech, missed signals, fatigue, inexperience.
This is why red-team thinking matters. Stress-test the plan like an adversary:
Where does information fall apart?
What if the client doesn’t disclose SI until 2 minutes before the next client?
What if a clinician freezes?
What if the EHR crashes? (Not that it would ever happen.)
What if “they would never” just nevered?

If you don’t intentionally look for the failure points and outliers, the system will reveal them, and in healthcare, that can be catastrophic.
Better to pressure-test your model now than learn its limits in a crisis.

How have you red-teamed your models?