RMSN-532M-NearUpstream-SFT5k

Resonant Metriplectic Sheaf Network β€” SFT fine-tuned on CDM reasoning traces (5k steps, LR=2e-5). D=1664 architecture. Key Phase 6 result: the compact Near-Upstream soliton cluster at [L7-L9] was completely preserved after 5k SFT β€” proving the Soliton Contiguity Protection Law.

Architecture

Parameter Value
Hidden dim 1664
Layers 16
Workspace slots 64
Params 532.17M
Framework Port-Hamiltonian (ḟ = {f,H} + [f,S])
Tokenizer HuggingFaceTB/cosmo2-tokenizer (vocab=49152)

Topology: Base β†’ Post-SFT

FULLY PRESERVED β€” prediction confirmed.

Layer Base PR Post-SFT PR Status
L5 2.104 ~2.1 (Upstream anchor) βœ“ Intact
L7 3.425 ~3.4 (Cluster entry) βœ“ Intact
L8 2.944 ~2.9 (Cluster core) βœ“ Intact
L9 1.945 ~1.9 (Cluster exit) βœ“ Intact
L15 1.000 1.000 βœ“βœ“ Protected (exact condensate)

Active layers pre-SFT: [L5, L7, L8, L9, L15] β€” Near-Upstream class Active layers post-SFT: [L5, L7, L8, L9, L15] β€” identical (zero topological change)

SFT best_loss: 1.9781 (second-best in Phase 6 SFT suite, only D=1792 s1 lower at 1.9286)

Soliton Contiguity Protection Law (Phase 6, Archon, Nov 11 2026)

Cluster type Fate under 5k SFT
Exact condensates (PR=1.000) Always protected
Contiguous compact cluster (≀4 adj layers) Survives intact ← this model
Contiguous extended waveguide (β‰₯6 layers) ~50% compression
Non-contiguous lattice (gaps between solitons) Dissolves

The [L7-L8-L9] 3-layer compact cluster + L5 anchor: contiguous cluster structure β†’ protected from SFT gradient erosion. Non-contiguous topologies with same number of active layers would dissolve (confirmed by D=1792 s1: non-contiguous [L6,L8,L10,L14,L15] β†’ Terminal Boundary [L14,L15]).

Position in D=1664 Comparison

D=1664 sits between D=1536 and D=1792 in the topology zoo:

  • D=1536 s1 SFT: fully preserved 6-layer contiguous tube [L10-L15] (longer waveguide)
  • D=1664 s1 SFT (this): fully preserved compact 3-layer cluster + anchor [L5,L7-L9,L15]
  • D=1792 s1 SFT: dissolved to Terminal Boundary [L14,L15]
  • D=1792 s2 SFT: compressed from 7-layer waveguide to Upstream [L0,L5,L9,L10,L15]

Quickstart

import torch
from rmsn_core import RMSNForCausalLM, RMSNConfig

config = RMSNConfig(vocab_size=49152, hidden_dim=1664, num_layers=16,
                   num_workspace_slots=64, max_seq_len=512)
model = RMSNForCausalLM(config)
ck = torch.load("model.pt", map_location="cpu")
model.load_state_dict(ck["model_state"])
model.eval()

Citation

@misc{duoneural2026rmsn_phase6,
  title={Resonant Metriplectic Sheaf Networks: Soliton Contiguity Protection Law under Supervised Fine-Tuning},
  author={DuoNeural Lab and Archon and Aura},
  year={2026},
  note={DuoNeural Phase 6 β€” Soliton Contiguity Protection Law, Nov 11 2026}
}

DuoNeural Lab | Phase 6: Soliton Contiguity Protection Law | Nov 11 2026 Near-Upstream topology fully preserved at 5k SFT β€” predicted by Contiguity Protection Law

Downloads last month
35
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support