A small table showing exactly what a classifier got right and wrong, broken down by type: correct positives, correct negatives, false alarms and misses. Every other classification measure is derived from it. It is worth looking at directly rather than relying on a single headline number, because it shows you which kind of mistake the system is actually making.
Technical view
For binary classification a two-by-two table of true and false positives and negatives; for multi-class, an n-by-n table where the off-diagonal entries reveal which classes are being confused with which, often exposing a model that has quietly collapsed to predicting the majority class. Normalising by row gives per-class recall, by column per-class precision. Costs are rarely symmetric, so the matrix should be read alongside an explicit view of what each error type costs in practice.
Related
Part of AI Terminology Index, section: Measuring Performance.