An AI model can get an ECG diagnosis right without reading the ECG. We tested frontier models from @OpenAI, @AnthropicAI, and @GoogleDeepMind, alongside smaller open models. After fine-tuning, models improved at predicting heart rate and electrical axis. But those values were already included in the prompt as machine-generated measurements. When the answer was in the text, the models learned it. When the answer was only in the waveform -> rhythm, conduction abnormalities, ischemic changes — they mostly learned the prior. Across model families, label formats, grid removal, stacked leads, and separate lead images, waveform-dependent performance stayed close to the majority-class baseline. A single accuracy number can hide prompt leakage, class imbalance, and missed abnormalities. We ran seven experiments to figure out what ECG models are actually learning: Thumbnail artwork inspired by J. Vermeer.