A split-control criterion for few-shot adaptation in single-cell perturbation prediction
Few-shot adaptation can help or harm single-cell perturbation prediction, but conventional evaluation shares a control estimate between calibration responses and held-out targets. We show how methods carry control error into their predictions and derive its contribution to score inflation. Nested control-depth experiments support inverse-depth scaling of inner-product inflation; correlation inflation also depends on the control noise-to-signal ratio. Independent-control scoring changed the rank of 40 of 52 configuration-dataset entries. We introduce SafeDelta, a split-control criterion that uses calibration data to decide whether to adapt. Across six datasets, mean gated gain was non-negative under the original two-control evaluation. On a deep plate, fixed predictions and decisions retained positive mean gains under three additional independent control references. The same reproducibility principle identified a fibrosis programme that localises to scar in human liver. These results inform adaptation decisions and control-depth planning.