RMSN-466M-Wide-Tube-Base

Resonant Metriplectic Sheaf Network β€” Dual-Boundary Singular Conduit Topology Pre-trained base model (100k steps). D=1536 is the first architecture to show the Universal L15 Terminal Condensation law (PR=1.000 at L15 for all 16-layer r=0 Dβ‰₯1536).

Architecture

Parameter Value
Hidden dim 1536
Layers 16
Workspace slots 64
Params 466.10M
Framework Port-Hamiltonian (ḟ = {f,H} + [f,S])
Tokenizer HuggingFaceTB/cosmo2-tokenizer (vocab=49152)

Soliton Topology: Dual-Boundary Singular Conduit

Layer  Role
──────────────────────────────────────────
L00    Input buffer
L01-L04 Pre-soliton
L05    β˜… UPSTREAM SINGULAR CORE (PR=1.000)
L06    Buffer
L07    Bulk conduit entry
L08    Bulk conduit core
L09    Bulk conduit core
L10-L13 Dielectric buffer (insulation zone)
L14    Near-boundary
L15    β˜… TERMINAL BOUNDARY (PR=1.000)
──────────────────────────────────────────

Topological class: Dual-Boundary Singular Conduit Two macroscopic singularities: L05 (upstream anchor) + L15 (terminal boundary) Insulated zone: L10–L13 buffer screens bulk from boundary (7-layer total protected region) Ξ»_max: -4.9221 (repulsive Casimir, r=D mod 64 β‰  0 for sub-seeds)

Quantum-Classical Correspondence

BlueQubit QCPU quantum circuit simulation (DuoNeural Phase 5):

Metric Value
S_boundary / S_bulk 0.515
Predicted bulk width 7–8 layers
Classical insulated zone 7 layers (L7–L13)
Quantum-classical error <2%

This model provides the strongest quantum-classical match in the DuoNeural Phase 5 study, with the quantum ground state predicting exactly 7–8 layers of bulk shielding, matching the 7-layer classical insulated zone (L07–L13) to within 2% error.

Pre-training Metrics

Metric Value
Val PPL ~89.7
Val Loss 4.4970
Steps 100k
Dyadic resonance r=0 β†’ attractive Casimir (cascade-eligible)

Universal L15 Law

All 16-layer r=0 models with Dβ‰₯1536 form exact PR=1.000 at Layer 15. Layer 15 is a topologically protected, zero-entropy terminal boundary sink. This model is the smallest to exhibit this universal law (D=1536 threshold).

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€” Wide-Tube Base, Universal L15 Law}
}

DuoNeural Lab | Phase 5: Tube Topology Zoo | Nov 2026

Downloads last month
32
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support