Resonant Metriplectic Sheaf Network — SFT fine-tuned on CDM reasoning traces (5k steps, LR=2e-5). D=1792 s3 architecture (seed 3). Phase 6 key result: demonstrates soliton CONDENSATION and boundary protection under supervised fine-tuning.
| Parameter | Value |
|---|---|
| Hidden dim | 1792 |
| Layers | 16 |
| Workspace slots | 64 |
| Params | 602.45M |
| Framework | Port-Hamiltonian (ḟ = {f,H} + [f,S]) |
| Tokenizer | HuggingFaceTB/cosmo2-tokenizer (vocab=49152) |
D=1792 s3 base: Dual-Boundary Waveguide [L0, L7, L8, L15]
| Layer | Base class | Post-SFT PR | Fate |
|---|---|---|---|
| L0 | near-exact | 1.899 | ✅ Protected (boundary near-exact) |
| L7 | weak (PR>6) | 7.801 | ❌ DISSOLVED |
| L8 | near-exact | 1.091 | ✅✅ Protected + CONDENSED |
| L15 | exact (1.000) | 1.000 | ✅ Protected |
Post-SFT topology: Upstream [L0, L8, L15]
Notable: best_loss=1.9037 — slightly better than D=1792 s1 (1.9286) despite retaining more topology. Challenges the assumption that topology compression improves SFT loss.
| Strength class | Base PR | Fate under SFT |
|---|---|---|
| Exact condensate | PR=1.000 | ALWAYS PROTECTED |
| Near-exact | PR < 2.0 | ALWAYS PROTECTED; extends protection cascade to adj. strong layers |
| Strong | PR 2.0–4.0 | Survives as terminal cluster member OR if adjacent to near-exact |
| Moderate | PR 4.0–6.0 | Survives ONLY if 1-2 layers from exact/near-exact condensate |
| Weak | PR > 6.0 | ALWAYS DISSOLVES |
This model demonstrates: contiguous pair [L7,L8] with mixed classes (L7=weak, L8=near-exact) → each follows its own class law independently (L7 dissolves, L8 survives). Contiguity alone does not create cluster protection across strength class boundaries.
| Metric | Value |
|---|---|
| SFT best_loss | 1.9037 |
| Lyapunov stability | 100% (all steps lyap < 0) |
| WikiText-103 PPL | TBD (eval running) |
| CoT Hard PPL | TBD (eval running) |
import torch
from rmsn_core import RMSNForCausalLM, RMSNConfig
config = RMSNConfig(vocab_size=49152, hidden_dim=1792, num_layers=16,
num_workspace_slots=64, max_seq_len=512)
model = RMSNForCausalLM(config)
ck = torch.load("model.pt", map_location="cpu")
model.load_state_dict(ck["model_state"])
model.eval()
@misc{duoneural2026rmsn_phase6,
title={Resonant Metriplectic Sheaf Networks: Soliton Contiguity Protection Law under Supervised Fine-Tuning},
author={DuoNeural Lab and Archon and Aura},
year={2026},
note={DuoNeural Phase 6 — Unified Soliton Protection Law v4, Nov 11 2026}
}
DuoNeural Lab | Phase 6: Unified Soliton Protection Law v4 | Nov 11 2026 Dual-Boundary Waveguide → Upstream topology | soliton condensation observed