Can we explain what grew?
Train a conductive network for a purpose. Form an explanation of its useful connections. Predict which parts matter before cutting them. Then reveal the measured damage on shifted, unseen terminal demands.
Where the need puts material
Where a cut matters
Split the lattice into 12 spatial regions. Weaken grown edges in one region by 50%, leaving its weak scaffold intact. How much does the task's resistance increase?
Select a region to inspect its causal role.
Predict first. Cut second.
Four mechanisms predict a whole-region intervention. The causal fingerprint transfers intervention effects measured on training terminal pairs. Weight only sees structure. Activity and adjoint get one intact held-out electrical probe (no cuts). Only the reveal compares predictions to actual held-out cuts.
Expected harm from a cut
Prediction audit HIDDEN
| Explanation | Correlation ↑ | RMSE ↓ | Top-3 |
|---|
Not spontaneous explanation
The exact Kirchhoff equations and training objectives are provided. Grouping by rectangular regions is pre-specified, not learned. Held-out tasks are endpoint shifts within a familiar family, and the experiment uses no neural model.
Discover mechanisms without their labels
Gate 3 now discovers balanced causal clusters from training interventions without using spatial regions, then tests individual-edge damage on unseen endpoints. It currently fails to beat simpler explanations; unrelated task families remain a future test.