RMSN-756M-Triple-Condensate-Base
Resonant Metriplectic Sheaf Network β Input-Early-Terminal Triple Condensate Topology Pre-trained base model (100k steps). D=2048 is the largest RMSN architecture in the Phase 5 topology study and the first to exhibit input boundary exact condensation (L0=1.000).
Key scientific finding: D=2048 is the first evidence of Topological Overcrowding β three exact singularities (PR=1.000) at L0, L6, and L15 freeze so much network capacity that val_ppl REGRESSES relative to D=1792 (84.62 vs 57.4), breaking the monotone D-scaling law observed across D=944β1792.
Architecture
| Parameter | Value |
|---|---|
| Hidden dim | 2048 |
| Layers | 16 |
| Workspace slots | 64 |
| Params | 755.57M |
| Framework | Port-Hamiltonian (αΈ = {f,H} + [f,S]) |
| Tokenizer | HuggingFaceTB/cosmo2-tokenizer (vocab=49152) |
Soliton Topology: Input-Early-Terminal Triple Condensate (NEW CLASS)
Layer PR Role
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
L00 1.000 β
β
β
INPUT BOUNDARY EXACT CONDENSATE (unprecedented)
L01 17.319 Buffer
L02 21.258 Buffer
L03 23.519 Buffer
L04 49.982 Buffer
L05 4.613 β
UPSTREAM SOLITON
L06 1.000 β
β
β
EARLY INTERIOR EXACT CONDENSATE (new position)
L07 27.039 Buffer
L08 26.641 Buffer
L09 52.877 Buffer
L10 1.697 β
MID-NETWORK SOLITON
L11 63.734 Dielectric buffer (high PR)
L12 62.007 Dielectric buffer
L13 61.075 Dielectric buffer
L14 58.224 Dielectric buffer
L15 1.000 β
β
β
TERMINAL BOUNDARY EXACT CONDENSATE (confirmed universal Dβ₯1792)
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
Active soliton layers: [0, 5, 6, 10, 15]
THREE exact singularities: L00=1.000, L06=1.000, L15=1.000
Topological class: Input-Early-Terminal Triple Condensate
- Unprecedented: L0=1.000 is the first input-boundary exact condensate in any RMSN base model
- New position: L6=1.000 is a new condensate position not seen in Dβ€1792
- Confirmed: L15=1.000 (universal for Dβ₯1792, holds at D=2048)
Topological Overcrowding Discovery
This model is the first evidence of Topological Overcrowding in RMSN architectures:
| D | Val PPL | Exact Singularities | Input (L0) |
|---|---|---|---|
| 1024 | 69.9 | 1 (L13) | Inert |
| 1280 | 65.4 | 2 (L8, L9) | Inert |
| 1536 | 60.3 | 2 (L5, L15) | Inert |
| 1792 | 57.4 π | 1β2 (varies) | Solitonic only |
| 2048 (this) | 84.62 π΄ | 3 (L0, L6, L15) | Frozen (1.000) |
Mechanism: Each PR=1.000 layer is a zero-entropy condensate that "freezes" the representation at that layer, reducing effective network capacity. At D=2048, three frozen layers (including the input L0) reduce information flow so severely that val_ppl regresses despite 25% more parameters.
Implication: There exists an optimal D range (~1664β1792) for the RMSN architecture at this training scale, beyond which dyadic resonance creates too many condensates.
Pre-training Metrics
| Metric | Value |
|---|---|
| Val PPL | 84.62 π΄ (D-scaling regression) |
| Val Loss | 4.4382 |
| Steps | 100k |
| D-scaling position | D=944(73.4)βD=1792(57.4)βD=2048(84.62) β REGRESSION |
| Optimal D predicted | ~1664β1792 |
Topological Bulk Shield Law
BlueQubit QCPU quantum circuit simulation (DuoNeural Phase 5):
| Metric | Value |
|---|---|
| S_boundary / S_bulk | 0.426 |
| Trend | 944(0.739)β1024(0.632)β1280(0.569)β1536(0.515)β1792(0.46)β2048(0.426) |
The quantum S_ratio continues its monotone decrease even though classical val_ppl regresses, suggesting the quantum-classical correspondence breaks down for the language modeling metric at topological overcrowding, while the topological structure remains classically predictable.
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 5d β Topological Overcrowding Discovery, D=2048 Triple Condensate}
}
DuoNeural Lab | Phase 5d: Topological Overcrowding | Nov 2026 First input-boundary exact condensate (L0=1.000) in RMSN base training D-scaling law breaks at D=2048: optimal D β 1664β1792
- Downloads last month
- 33