Evaluation frame
The useful question is not only whether a model predicts held-out data. It is whether it can reduce experimental ambiguity.
Next pass
- Define perturbation classes.
- List benchmark limitations.
- Track examples where predictions suggest a real experiment.