anima-deep-convmoe-L8 (H_1584 depth-RF probe · NEGATIVE result)

Deep ConvMoE L=8 (RF-expanded, d=3784, 3 experts, SLW) trained on the 4-cell corpus (ko/en × general/sns · steps 2000 · seq-len 1024). Tests whether receptive-field expansion (L4→L8) cracks the G1 recombination wall.

• Training: CLEAN — loss 5.636→1.552 DESCENT · registers_DESCENT 4/4 · val_CE(pooled) 1.71 ≪ uniform 5.545 · clm_decodable=True.
• G1 (ρ·weave) verdict: 🧱 FAIL — best_distinct=1 (need ≥2 & >max_single) — SAME floor as production L4.
• Conclusion: depth/RF expansion does NOT crack G1 recombination at engine-native 303M-class byte-LM (H_1584 FALSIFIED). Consistent with the DPI meta-law: an additive-solvable loss escapes to the additive floor regardless of capacity/RF. The numpy reachability probe (conv_L8 reach 1.47e-3 REACHABLE) proved information flow exists but the trained model does not learn to use it for recombination.

sha256: 7d221ec81cee1543de012f01ba2d060bd1f304bd5ed9fab865f27c3b2e5af178