External-transfer benchmark
External-transfer diagnostic
55-target aggregate transfer remains unsupported
The metric adapter executes on an external aggregate-delta object. It is useful for method diagnostics but is neither cell-distribution evaluation nor a VCC submission score.
Multi-metric run on 55 external aggregate perturbation targets
MAE-delta, delta correlation and L1-PDS now run beside the same ridge, training-mean and zero-change predictions. The source endpoint remains NO_DIRECTIONAL_TRANSFER_SUPPORT.
| Method | RMSE Δ | MAE Δ | PDS L1 | Pearson Δ | Spearman Δ | Effect-size ρ |
|---|---|---|---|---|---|---|
| Safe ridge transferdiagnostic model | 1.0928 | 0.1861 | 0.5147 | -0.0148 | -0.0013 | -0.0554 |
| Training-response meanbaseline | 1.0925 | 0.1844 | 0.5134 | -0.0161 | -0.0027 | N/D |
| Zero changebaseline | 1.0896 | 0.1707 | 0.5134 | N/D | N/D | N/D |
N/D means not defined. A constant or zero-variance prediction cannot support a correlation estimate. N/D is not zero and is excluded from metric comparisons.
Truth-norm matched: 0.5431 · median factor 14.52×. This uses paired observed norms and is sensitivity-only, never a fair hidden-test score.
What changed. The metric engine is operational on aggregate deltas. DES/DEG recovery and cell-distribution metrics remain locked because this prediction object has no genuine cell-level distribution.