Does a network learn a need,
or memorize its endpoints?
Specialist vs diverse growth, same three electrical solves per update. Choose a checkpoint from validation before testing unseen demands. Then lock causal predictions and reveal how the network really responds to cuts.
Choose before peeking
The minimum of the three-case validation resistance selects one checkpoint in each arm. No hidden test values enter this chart.
| Growth rule | Selected step | Selected validation ↓ | Step-320 validation ↓ | Material |
|---|
Two ways of growing
Same graph, initial conductance, material cap (24), learning rate and training solve count; diverse growth covers a larger set of endpoints with different examples per step. Checkpoint selection uses validation costs, not test costs.
Say which regions matter before cutting them
Each of four models uses only the three original training demands. A non-cryptographic state digest locks its predictions to the selected conductances. Cuts halve grown conductance in one of 12 spatial regions; the base scaffold stays intact. White dots appear only after reveal.
What survived new demands?
The 24-pair sweep was generated deterministically from unseen endpoints, disjoint from training, augmentation and validation. Three separately listed stress pairs offer a harsher shift. Lower resistance is better. The uniform control uses the same actual grown material.
| Arm | Selected sweep ↓ | Final sweep ↓ | Uniform sweep ↓ | Selected stress ↓ |
|---|
Causal intervention transfer
| Arm | Blind predictor | Correlation ↑ | RMSE ↓ | Top 3 |
|---|
Correlation is descriptive across 12 predefined regions, not an independent-sample confidence interval. Both test sets are fully visible in the source code and this is procedural, not cryptographic blinding.
When simplicity wins
Early stopping and diversity can help, but aren't guaranteed to beat an equal-material uniform network on unseen endpoints. A causal fingerprint isn't better merely because it involved interventions: test it against material and training-flow baselines.
Known physics and region labels
Fixed rectangular region boundaries, prescribed Kirchhoff dynamics, supervised demands and deterministic runs. Nothing here proves Goodfire-style clean semantic circuits, neuron birth or transferable mechanisms across unrelated task families.