The Confessions of a Protein Language Model: Steering, Activations, and What Predicted Confidence Cannot Certify
Protein generate-and-screen pipelines use interventions and readouts to select candidates, but rarely test whether these track their intended targets. We a