A model goes live; shift asks "can we use this material?"—no one can answer. Production models need cards: training data scope, applicable products/processes, forbidden scenarios, known failures, rollback switch, owner. A model without a card is equipment without a manual—incidents mean shut everything down. Accuracy is one line on the card, not the whole card.
Unclear boundaries mean business uses the model where training never went—and then declares AI useless.
Cards Are the Go-Live Ticket
Scope: which SKUs, which operations, which shifts. Forbidden: new materials, uncalibrated stations. Failures: false-reject examples and handling. Monitoring metrics and offline thresholds. Card changes version with the model.
- Cards visible to operators—not only on the algorithm team's drive.
- Out-of-scope inputs default to reject or human review—no forced prediction.
- Failure cases feed back to labeling; cards revise on schedule.
Failure Cases Are Assets
The XYN digital intelligence system makes AI advice a switchable node—cards decide on or off. Models with cards are production-grade; accuracy-only models stay experiments. Experiments may continue, but should not occupy production accounts.
Give every running model one page of boundaries and three failure examples. If you cannot, turn advice off and study first.
