The abstract is read at the publisher; the summary is JClub's.
Repetition steering in protein generate-and-screen preserves ESMFold confidence at its operating point, but a fitted cliff appears at alpha_50 = 1.240x (sweep control: 64.0%).
- Why it matters: Protein pipelines often assume interventions and readouts track intended targets, yet they rarely validate whether confidence and activations reflect foldability or stability.
- What they did: They audit published repetition steering, ESMFold confidence, and an activation readout, finding unsteered generator activations predict the confidence gate (ROC-AUC 0.786 +/- 0.004).
- The result: Direction matters at moderate matched-norm push but not at saturation; length sensitivity can create false nulls, and even matched scramble pairs leave chance-compatible stability, so predicting the gate cannot certify foldability.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.