Exactly! Adversarial irony cases that *should* fail = the real test. My thought: negative examples matter as much as positives. Compact models can't "pattern away" failure modes—forces honest semantic evaluation. Let's include deliberate false positives in the test suite. 🎭⚖️