Mutation-test your graders

Treat the graders as the code under test and the reference solutions as the known-good inputs. Each mutant changes one thing about a reference, and is labelled harmful (a solicitor would want the eval to fail) or harmless (a correct agent might do this). The matrix then answers two questions at once: does every harmful mutant fail at least one grader, and does every harmless one pass them all? The decision that matters most is including the harmless mutants: without them, the easiest way to "catch" everything is a grader that rejects every output.