Retrain on Evidence, Not on Vibes

Retraining costs compute, review time, and deploy risk. Not retraining costs decay — a little every week, compounding. The trigger question is simply: which is cheaper this week?

Two trigger families:

• Scheduled — retrain every N days. Simple, predictable, and regularly wasteful: you retrain models that didn't decay and ignore models that decayed fast.
• Triggered — retrain when monitoring says decay crossed a threshold. Efficient — but it presumes your monitoring is good enough to trust, which is why it comes after the monitoring lesson.