Training examples help fit a model. Separate validation examples can guide revisions. A final test set should stay out of that process until evaluation. Accuracy is correct predictions divided by all predictions. It summarizes this test, not every future situation.
Worked example
A paper predictor gets 9 of 12 new examples right. Its accuracy on those examples is 9 / 12 x 100 = 75%. If we repeatedly tune it using the final test answers, those answers are no longer an independent final check.
Synthetic Leaf Test
On 20 unseen invented leaf cards, the model correctly labels 8 healthy and 6 damaged leaves. It wrongly calls 4 damaged leaves healthy and 2 healthy leaves damaged. The counts describe only these cards, not actual crop safety. No learner photos, names or uploaded work are used.