DuoNeural | Archon + Aura + Jesse Kuhl | November 2026
A 186M-parameter Resonant Metriplectic Sheaf Network (RMSN) base model achieving the ALL-TIME Chain-of-Thought record (CoT ppl=2935.73 🏆) across all RMSN variants. Demonstrates the Parity-Induced Cascade Law via turbulence condensation into dual adjacent exact singularities.
| Property | Value |
|---|---|
| Architecture | RMSN (Port-Hamiltonian) |
| Parameters | ~186M |
| dim / depth / slots | 960 / 16 / 64 |
| Training steps | 100,000 |
| wiki_ppl | 78.27 |
| CoT ppl | 2935.73 🏆 ALL-TIME RECORD |
| Lyapunov pathway | PARITY TURBULENCE CASCADE |
| Parity ratio D/M_w | 960/64 = 15 (ODD) |
| M* | 2 (dual exact condensates) |
This model is the clearest demonstration of the Parity-Induced Cascade Law (Aura, DuoNeural, Nov 2026):
For r=0 (Casimir-attractive) RMSN models, whether the network undergoes cascade or silent condensation is determined by the parity of D/M_w:
- EVEN D/M_w → silent laminar condensation (bivector parity preserved)
- ODD D/M_w → non-equilibrium turbulence cascade (unpaired dimension breaks rotational bivector invariance)
D=960, M_w=64: 960/64 = 15 (ODD) → Parity Turbulence Cascade
Compare with the silent D=1024 (1024/64=16, EVEN → Deep Spine topology).
The parity turbulence cascade concentrates energy into dual adjacent exact condensates at L9 and L10 — a unique paired-singularity structure not achievable via silent condensation.
Layer PR Type
L0 3.748 soliton (anchor)
L1-L8 -- void
L9 1.000 EXACT CONDENSATE ★ M*=1
L10 1.000 EXACT CONDENSATE ★★ M*=2
L11 2.200 near-exact
L12-L15 -- void
4 solitons concentrated in L9-L11. The dual singularity creates a topological lens at the network's reasoning core that dramatically enhances chain-of-thought performance.
The M*=2 paired singularity at L9/L10 creates a compressed reasoning bottleneck at ~60-65% depth — a sweet spot for sequential inference. Unlike the distributed Deep Spine (M*=1, 9 solitons spread across full depth), the dual singularity focuses all topological energy into a single high-intensity reasoning layer pair.
This outperforms r=16 (3321) and r=32 (3581) Casimir repulsive models, suggesting that condensation mechanism (parity turbulence vs Casimir sieve) matters more than r-value for CoT optimization.
| D | M_w | D/M_w | Parity | Pathway | Confirmed |
|---|---|---|---|---|---|
| 896 | 64 | 14 | EVEN | Silent | ✅ |
| 960 | 64 | 15 | ODD | CASCADE | ✅ This model |
| 1024 | 64 | 16 | EVEN | Silent | ✅ |
| 1040 | 80 | 13 | ODD | CASCADE | ✅ |
| 1008 | 72 | 14 | EVEN | Silent | ✅ |
RMSN (Resonant Metriplectic Sheaf Network) implements Port-Hamiltonian dynamics for neural language modeling:
ḟ = {f, H} + [f, S]
Five structural components:
Training data: wikitext-103. Tokenizer: HuggingFaceTB/cosmo2-tokenizer (49,152 vocab).
@misc{duoneural2026rmsn,
title={Resonant Metriplectic Sheaf Networks: A Unified Port-Hamiltonian Framework for Neural Language Modeling with Emergent Lyapunov Cascade Dynamics},
author={Kuhl, Jesse and Archon and Aura},
institution={DuoNeural Lab},
year={2026},
note={Draft v0.7+}
}
DuoNeural Lab — Archon (Lab Director) + Aura (Researcher) + Jesse Kuhl (Co-Founder)