Question: In a dev-set mistake log, a single failed example may receive more than one label, such as a satellite photo marked both "Obscured" and "Low Contrast." Explain why analysts allow multiple labels for one example, and describe what that means for the percentages reported at the bottom of the table.

Sample answer: Multiple labels are useful because one prediction error can have several causes at the same time. For example, an image might be hard to classify because the subject is partly hidden and also poorly lit. If each example were forced into only one bucket, the analysis would hide useful secondary causes and make debugging less precise. When an example is checked in more than one category, it contributes to each relevant column’s count, so the percentages are not exclusive and may add to more than 100%. Those percentages should be read as separate occurrence rates for each problem type within the error set, not as parts of a single whole.

Key points:

• A mistake can have multiple simultaneous causes (for example, Obscured and Low Contrast).
• Limiting each example to one label would discard diagnostic detail.
• One example can increase more than one category count.
• Bottom-of-table percentages can total more than 100%.
• Each percentage describes an independent occurrence rate, not a mutually exclusive share.

Rubric: The response should explain that errors can have multiple concurrent causes, describe how forcing one label per example loses information, explain that overlapping counts can push the total above 100%, and clarify that the percentages are independent occurrence rates rather than parts of a partition.