model card: checkpoints + ablation

README.md

@@ -19,3 +19,14 @@ Graph files (`edges.idx.bin`, `indptr.bin`, `input_idx.bin`, `readout_idx.bin`,
 identical to [naklitechie/flymate-chess](https://huggingface.co/naklitechie/flymate-chess) and are
 fetched from there. Play: https://flymate.naklitechie.com/go · code: https://github.com/NakliTechie/flymate
 · training: gofly run 4. FlyWire's non-commercial terms apply to anything derived from the graph.
+
+## Torch checkpoints and the ablation (`checkpoints/`, `ablation/`)
+
+| file | wiring | held-out KataGo top-1 · ownership · value MAE |
+|---|---|---|
+| `run4-real-30.1.pt` | FlyWire v783 as published | **30.1%** · 74.9% · 0.432 — source of the browser bundle above |
+| `run5-shuffled-30.5.pt` | **degree-preserving random rewiring** (`ablation/connectome-shuffled-seed0.npz`) | **30.5%** · 75.2% · 0.430 |
+
+Run 5 is run 4 with one change: a configuration-model shuffle that keeps every neuron's in/out degree, every synapse's sign and count, and 0.61% of the original edges. Same data order, seed, hyperparameters, 20k steps. The lead changed hands 8 times over 20 evaluations; the shuffle finished +0.5 pt ahead. On 9×9 Go the specific wiring diagram contributes nothing measurable — what the network uses is the graph's statistics (degree distribution, sign balance, sparsity), i.e. it is a good random sparse reservoir. Scope: one task, one board size, one shuffle model; a shuffle that also destroys degrees and signs is the next control.
+
+`connectome-shuffled-seed0.npz` has the same CSR layout as the real graph (`indptr`, `indices`, `sign`, `count`) and is required to run `run5-shuffled-30.5.pt`. FlyWire's non-commercial terms apply to it as a derivative of the graph.