Grouping failures into modes

After lesson 4 you have a stack of one-sentence descriptions. This lesson turns them into categories. Read them together and group the ones that describe the same underlying problem, then give each group a short name.

A failure mode is a named, countable category of failure. "Confuses returns with billing when money is mentioned" is a failure mode. "The model is bad at returns" is not, because you cannot count it or tell whether it has improved.

• Name the behaviour, not the model. The mode should still make sense after you switch models.
• Keep the set small. Five to eight modes covers most of what a single agent does wrong.
• Count each one. The counts tell you what to work on first.
• Keep a mode called "label disagreement" for cases where the dataset is at fault.

The counts are the point. Once every failure carries a mode, you can say "forty percent of our errors are one confusion between two categories" — which is a specific, fixable statement. It also gives you a way to check later whether a change helped: the mode you targeted should shrink, and the others should hold steady.