Accuracy is a starting point, not a conclusion

One number, several unanswered questions

A model reaches 90% accuracy. That sounds promising, but it leaves several questions open. Which examples did it miss? Were some classes much harder than others? Would the errors be acceptable in the setting where the model will be used?

An overall score compresses a set of predictions into one number. That compression is useful for comparison. It also removes information.

Start with the confusion matrix

For a classification task, a confusion matrix is a practical next step. It shows how the reference classes and predicted classes line up. A model may perform well on a common class while repeatedly confusing a less common one.

In a binary task, distinguish false positives from false negatives. The two errors may have different consequences. A decision that creates an extra review task is different from one that misses an example that needed attention.

That does not make accuracy useless. It means accuracy answers a narrower question than we sometimes ask it to answer.

Inspect the examples behind the counts

A confusion matrix tells you where errors accumulate. Inspecting the examples can help you form explanations worth testing.

For an image classifier, consider a short error-review routine:

Inspect a few mistakes from each confused class pair.
Look for changes in lighting, framing, background, or image quality.
Check whether the labels are clear and consistent.
Compare training and evaluation preprocessing.
Write one testable hypothesis before making a change.

These observations are clues. They do not establish a cause on their own. A background pattern, for example, may suggest a shortcut, but that explanation needs a controlled test.

Separate a useful change from a lucky result

After an improvement, ask whether the comparison was fair. Did the data split stay the same? Did other settings change at the same time? Could variation between training runs explain part of the difference?

Keep a simple experiment log. Record the data version, split, configuration, seed, metrics, and an observation about the mistakes. The log is often more useful than another unlabelled checkpoint.

Let the decision context guide the next question

The most useful next metric depends on the task. Precision and recall help describe different aspects of positive predictions. Threshold analysis helps when a score must become an action. Class-level results help reveal uneven behavior.

A good evaluation does not need every available metric. It needs enough evidence to support the claim you are making.

Try the decision threshold lab to see how identical scores can lead to different errors. For a reusable review routine, use the model evaluation checklist.