I'mn talking about math research. Of course you can test things: does the formalized proof check? This is what makes automated math such an attractive target for AI.