I still remember giving Mathematica a relational equation of the atomic radius expectation values for it to integrate by parts and collect terms... at the time it failed to find the right integrating factor and gave gibberish. Probably works now.
The midterm was open book, he even shared the previous year's exams. The final, as expected, closed book, thank you very much.
That's also how you can tell if your calculator is "lying" to you (or you typed wrong). I guess I have a few similar tools in spaces where I'm more experienced to see when "hallucinations" and gibberish are being generated by LLMs. Of course having it "check sources" and evaluate its own solutions sometimes also works, if those are reliable, but you're on the hairy edge at that point.
Number sense is such a nebulous concept, but it captures what I've seen people struggle with. A general sixth sense about where an equation should or shouldn't go, a feeling that random numbers don't appear quite random enough and might have a pattern, or even recognizing that having a pattern signifies some relationship that otherwise has no evidence.