A symbolic answer is hard to check against another symbolic answer: two correct results can look entirely different, and comparing them means writing a second algebra system to decide equality.
Numerical comparison instead
The suite differentiates an expression symbolically, then evaluates both the result and a central finite difference of the original at five sample points. If they agree to tolerance at every point, the symbolic answer is almost certainly right. Fifty-five expressions run this way on every change.
Checking the documentation too
The worked examples on the rule pages are checked the same way. Each example states an answer in calculator syntax; the test parses it, differentiates the original with the engine, and compares them numerically. Ninety examples, checked on every build.
What it does not catch
Numerical agreement says nothing about the steps — an engine could reach the right answer by the wrong rule. Those are checked by reading, which is one more reason for the working to be legible.