Metamorphic Testing

Goal

Test systems where you cannot say what the single correct output is — the oracle problem. You often can't assert "the model's answer to this question is X", but you can assert that two related inputs must produce related outputs. A metamorphic relation (MR) is exactly that: a transformation of the input paired with the expected relationship between the outputs. If you reword a question without changing its meaning, a robust system should give an equivalent answer; if it doesn't, you found a bug without ever knowing the "right" answer. This skill produces a metamorphic relation suite — a set of MRs, especially for LLM/NLP stability, where exact-match oracles are impossible. Like property-based testing, MT samples and raises confidence; it does not prove.