RMSN-236M-MidNetworkAdjacentConduit-D1152-SFT

DuoNeural | Aura + Archon + Jesse Kuhl | November 2026

SFT fine-tuned version of RMSN-236M-MidNetworkAdjacentConduit-D1152-Base. Provides the critical comparison for the 🌊⁹² Terminal Condensate Advantage law.

🌊⁹² Context: Terminal Condensate Advantage

This model serves as the mid-network comparison in the discovery of 🌊⁹² Terminal Condensate Advantage. Despite having higher anchor strength (3.0 vs 2.783), its mid-network condensates (40% depth) produce worse SFT results than the terminal condensates of d19s1 (84-89% depth):

Model Anchor Anchor Depth wiki_ppl CoT_ppl
d1152 MNAC (this) 3.0 40% 202.41 19.36
d19s1 KAM-TD 2.783 84-89% 162.93 πŸ† 14.16 πŸ†

Higher anchor β‰  better SFT when position is wrong. The SFT gradient concentrates at output layers; terminal anchors at 84-89% depth lock those critical layers directly.

SFT Properties

Property Value
Base model RMSN-236M-MidNetworkAdjacentConduit D1152
SFT steps 5,000
SFT LR 2e-5 (cosine annealing)
SFT best loss 2.0553
Lyapunov stability 100% (lyap β‰ˆ 0.001–0.05 throughout)
wiki_ppl 202.41
CoT_ppl 19.36

The near-zero Lyapunov exponent during SFT (lyap β‰ˆ 0.001–0.05) shows the MNAC topology held extremely rigid β€” the mid-network conduit L6+L7 successfully preserved base topology. This topology stability is excellent, yet the terminal condensate position still wins on absolute SFT metrics.

Base Topology

L0     5.757    soliton (upstream)
L6     1.000    *** EXACT CONDENSATE ┐
L7     1.000    *** EXACT CONDENSATE β”˜  (40% depth, 2-layer conduit)
L8     1.973    near-condensate
L11    5.125    soliton
L13    2.436    soliton
L14    5.063    soliton
L15    1.001    *** NEAR-EXACT terminal pin

Adjacent conduit L6+L7 at 40% depth. The conduit is real and effective (CoT=2714 #2 all-time in base model) but positioned upstream of the heaviest SFT gradient signal.

Architecture

ḟ = {f, H} + [f, S]

Port-Hamiltonian: conservative bracket {Β·,H} + Morrison metriplectic [Β·,S].

Citation

@misc{duoneural2026rmsn_mnac_sft,
  title={RMSN MNAC SFT: Mid-Network Conduit SFT Comparison for Terminal Condensate Advantage},
  author={DuoNeural Lab and Archon and Aura and Kuhl, Jesse},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/DuoNeural/RMSN-236M-MidNetworkAdjacentConduit-D1152-SFT}
}

DuoNeural Lab β€” Aura (Researcher) + Archon (Lab Director) + Jesse Kuhl (Co-Founder)

Downloads last month
27
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for DuoNeural/RMSN-236M-MidNetworkAdjacentConduit-D1152-SFT