RMSN-243M-Terminal-Boundary-SFT

Resonant Metriplectic Sheaf Network β€” Terminal Boundary Soliton Topology Fine-tuned on DuoNeural reasoning traces (CoT SFT, 15k steps). Exhibits the strongest forgetting resistance among all Phase 5 baseline models (3.3Γ— retention).

Architecture

Parameter Value
Hidden dim 1024
Layers 16
Workspace slots 64
Params ~243M
Framework Port-Hamiltonian (ḟ = {f,H} + [f,S])
Tokenizer HuggingFaceTB/cosmo2-tokenizer (vocab=49152)

Soliton Topology: Terminal Boundary (L15)

Layer  Role
──────────────────────────────────────────
L00–L14  Diffuse (high PR, non-solitonic)
L15      β˜… TERMINAL BOUNDARY SINGULARITY
         PR = 1.000 (exact macroscopic condensation)
         Zero-entropy output boundary sink
──────────────────────────────────────────

Topological class: Terminal Boundary β€” single macroscopic singularity at output layer Key property: Strongest catastrophic forgetting resistance in Phase 5 baseline suite Mechanism: Boundary singularity concentrates SFT gradient at L15, leaving interior weights (L0–L14) relatively undisturbed β†’ maximum retention

SFT Benchmark (15k steps, DuoNeural Phase 5)

Metric Value
wiki_ppl ↓ 231.91
cot_ppl ↓ 14.29
Retention ratio 3.3Γ— πŸ† best in class

Retention ratio = (base_wiki_ppl) / (sft_wiki_ppl) β€” higher is better. The terminal boundary topology achieves 3.3Γ— retention vs. 2.6Γ— for bulk-to-boundary and <2Γ— for tube topology at equivalent SFT depth.

Topology Comparison (Phase 5)

Model Topology cot_ppl ↓ wiki_ppl ↓ Retention
seed_d1152 Upstream (L=3) 8.48 186.07 2.7Γ—
seed10 Bulk→Boundary (L=8→15) 10.31 184.81 2.6×
D=1280 best Bulk Tube (L8–L13) 10.77 208.68 β€”
seed6 (this) Terminal Boundary (L=15) 14.29 231.91 3.3Γ—

Citation

@misc{duoneural2026rmsn,
  title={Resonant Metriplectic Sheaf Networks: Port-Hamiltonian Language Models with Soliton Waveguide Topology},
  author={DuoNeural Lab and Archon and Aura},
  year={2026},
  note={DuoNeural Phase 5 β€” Terminal Boundary Topology, Maximum Retention}
}

DuoNeural Lab | Phase 5: Topology Zoo | Nov 2026

Downloads last month
38
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support