RMSN-532M-EarlyNucleation-SFT2500
Port-Hamiltonian Neural Language Model — DuoNeural Lab
D=1664 · 16 layers · 64 slots · 532M params
Overview
Instruction-tuned RMSN with Early Nucleation topology: SFT causes mid-layer solitons (L7/L8) to dissolve while nucleating a new front-anchor condensate at L5. This topology reorganization is architecture-driven, not LR-schedule-dependent.
Soliton Topology (Post-SFT 2500 steps)
| Layer | PR | Class |
|---|---|---|
| L5 | 2.764 | Near-Exact |
| L9 | 2.210 | Near-Exact |
| L15 | 1.000 | EXACT |
Base topology: Crystallized Cluster [L7, L8, L9, L15]
Post-SFT topology: Early Nucleation [L5, L9, L15] — L7/L8 dissolved, L5 nucleated
Evaluation
| Metric | Score |
|---|---|
| WikiText-103 | 158.77 |
| CoT Reasoning | 20.50 |
Best 2500-step SFT wiki retention.
Key Finding
L5 nucleation is architecture-intrinsic: identical topology produced with both standard and compressed LR schedules (confirmed by controlled comparison). SFT causes "condensation migration" from mid-layers to early layers under CoT supervision.
Architecture
RMSN (Resonant Metriplectic Sheaf Network) — Port-Hamiltonian dynamics: ḟ = {f,H} + [f,S] | Lyapunov stable | Phase-space conserving
Citation
DuoNeural Lab — Jesse + Archon + Aura | Nov 2026