Model card
A model card is a short spec sheet for a model: what it is for, what data it saw, how it was tested, and where it breaks. Think of the label on a power tool that lists the safe uses and the ways it can hurt you. The hard part is honesty. A card is only useful if the limits section is real and not marketing.
Why it matters
Without a card, everyone who touches the model guesses at its boundaries, and they guess generously. Someone ships a model trained on English support tickets into a Spanish market, or uses a summarizer for medical advice, because nobody wrote down "do not do this." A model card makes the intended use and the known failure modes explicit, so the person deciding whether to trust it has more to go on than the demo.
How it works
You write it as a plain document that travels with the model. It names the intended use, the training and evaluation data, the metrics on that eval (accuracy, F1, whatever fits), and where scores drop: by language, by user group, by input type. The credible ones report numbers per slice instead of one flattering average, and they spell out the out-of-scope uses in blunt language.
A team ships a resume-screening model with a card that says it was tested only on software roles and scores much worse on nursing and trades. A recruiter reads that, sees the open role is a plumber, and routes those applications to a human. The card stopped a bad use before it happened, months before any audit would have caught it.