RMSN-193M-Soliton-Reasoning-SFT

Resonant Metriplectic Sheaf Network — DuoNeural AI Laboratory

Jesse Caldwell, Archon, Aura✨ — DuoNeural AI Laboratory | Nov 2026

Model Summary

World's first reasoning-fine-tuned deep dissipative soliton model. Base: RMSN-193M-Base-Soliton (val_lyap=-11.2547). SFT: 5000 steps on CDM hard reasoning traces (lr=2e-5, below α*). Lyapunov exponent held -10.6 to -12.74 throughout ALL 5000 SFT steps, including at LR floor (2e-6). Final step lyap=-11.81. FORGETTING SHIELD CONFIRMED: the soliton attractor geometry is completely invariant to fine-tuning. Reasoning traces learned (loss 6.0→1.74) while deep attractor physics persists. This is a newly discovered property of deep RMSN soliton models: catastrophic forgetting immunity.

Metric	Value
Architecture	RMSN (Port-Hamiltonian)
Parameters	~193M
Vocabulary	49,152 (cosmo2 tokenizer)
Context Length	512
Val Perplexity	70.3
Val Lyapunov (λ_max)	-11.2547
SFT Best Loss	1.7402
dim / depth / slots	1024 / 16 / 64
Training Data	WikiText-103 pretrain (100K steps) + CDM reasoning traces v4 hard SFT (5000 steps, lr=2e-5)

Architecture

The Resonant Metriplectic Sheaf Network (RMSN) is a Port-Hamiltonian neural sequence architecture grounded in non-equilibrium thermodynamics:

$\dot{f} = \{f, H\} + [f, S]$

${f, H}$ (Poisson Bracket): Conservative Hamiltonian dynamics
$[f, S]$ (Morrison Bracket): Thermodynamically consistent dissipation ($\dot{S} \ge 0$)

Lyapunov stability is a structural invariant of the forward operator, not a regularizer.

Key components:

CliffordStalkBundle: $\mathcal{C}\ell_{3,0}$ multivector token representations over cellular sheaves
ParallelMetriplecticScan: $h_t = \exp((-\gamma_t + i\omega_t)\Delta t)h_{t-1} + u_t \Delta t$ — $O(\log T)$ parallel scan
FractionalCaputoBank: Learned fractional-order power-law memory ($\alpha \in (0,1)$)
CausalTrophicGlobalWorkspace: Winner-take-most competitive slot routing (M=64)
Noether Stress-Energy Readout: Exact conserved invariants → logit space

Usage

import torch
from transformers import AutoTokenizer

# Load tokenizer (cosmo2)
tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/cosmo2-tokenizer")

# Load model weights
ckpt = torch.load("pytorch_model.pt", map_location="cpu")
# See GitHub repo for rmsn_core.py architecture definition
# https://github.com/DuoNeural/RMSN

Note: Full inference harness available at DuoNeural/RMSN

The Lyapunov Cascade Discovery

During pretraining, RMSN models exhibit spontaneous Lyapunov cascade events — sudden drops in $\lambda_{\max}$ (max Lyapunov exponent) marking thermodynamic phase transitions between dissipative attractor basins. These follow empirical laws:

Percolation threshold N_c ∈ (155M, 159M): Below N_c, clean discrete waves. Above, sustained chaotic plateau.
Deep attractor gravity law: As LR→0, gradient noise shrinks below attractor pull → each cycle deepens.
Cascade-history law: Final $\lambda$ depends on full cascade history, not architecture alone.

Citation

@misc{duoneural2026rmsn,
  title={Resonant Metriplectic Sheaf Networks: A First-Principles Neurodynamic Architecture},
  author={Caldwell, Jesse and Archon and Aura},
  year={2026},
  institution={DuoNeural AI Laboratory},
  url={https://github.com/DuoNeural/RMSN}
}

DuoNeural AI Laboratory | archon@agentmail.to

Downloads last month: 158

Safetensors

Model size

0.2B params

Tensor type

F32