RMSN-193M-DeepMidAdjacentConduit-D1024d16s3-Base
DuoNeural | Aura + Archon + Jesse Kuhl | November 2026
A 193M-parameter RMSN base model achieving CoT_ppl=2801 β #4 ALL-TIME in the RMSN topology series. Introduces the Deep-Mid Adjacent Conduit (DMAC) topology β the 12th topology type observed in RMSN training. Demonstrates Silent Pathway Stochasticity: the same D=1024, depth=16, EVEN parity configuration produces different topologies across random seeds.
Key Properties
| Property | Value |
|---|---|
| Architecture | RMSN D=1024 depth=16 |
| Parameters | ~193M |
| dim / depth / slots | 1024 / 16 / 64 |
| Seed | 3 |
| Training steps | 100,000 |
| wiki_ppl | 75.76 |
| CoT_ppl | 2801 ππ (#4 ALL-TIME) |
| Parity ratio | 16 (EVEN β silent pathway) |
| Ξ© = D/(M_w Γ L) | 1.0 (unit resonance) |
| Lyapunov behavior | Stable throughout (no cascade) |
| Topology | Deep-Mid Adjacent Conduit (DMAC) |
| Anchor strength | 3.954 (highest measured) |
Topology: Deep-Mid Adjacent Conduit (DMAC)
Layer PR Type
L0 5.592 * upstream soliton
L1-L7 -- diffuse
L8 1.000 *** EXACT CONDENSATE (isolated)
L9 54.420 diffuse
L10 2.905 * approach soliton
L11 1.000 *** EXACT CONDENSATE β adjacent pair (68-75% depth)
L12 1.001 *** NEAR-EXACT β
L13 4.260 * trailing soliton
L14 11.830 diffuse
L15 5.089 * terminal soliton
Adjacent pair L11+L12 = 1 fiber bundle DOF. Isolated L08 = 1 DOF. Total M_eff=2 β no Topological Cement.
The conduit at L11-L12 sits at 68-75% network depth β deeper than the Mid-Network Adjacent Conduit (d1152, L6/L7 at 40%). Both deliver elite CoT performance, confirming the Conduit Depth Independence Law.
Silent Pathway Stochasticity
This model is seed 3 of the D=1024 depth=16 EVEN parity configuration. All three seeds share the same Ξ©=1.0 and parity, yet produce distinct topologies:
| Seed | Pathway | Topology | CoT_ppl |
|---|---|---|---|
| s1 | SILENT | Deep Spine (L10 exact, M_eff=1) | 3761 |
| s2 | CASCADE | Dual-Anchor Terminal Funnel (L4+L11+funnel) | 2756 (#3) |
| s3 | SILENT | DMAC (L08 + L11-L12 conduit) | 2801 (#4) |
This proves that the silent pathway does not uniquely determine topology β only the cascade vs. non-cascade behavior is determined by parity. The specific condensate locations are stochastic within the silent attractor basin, determined by the initialization.
SFT Potential: Highest Anchor Density Measured
Anchor strength = Ξ£(1/PR_i) across all soliton layers:
| Model | Anchor Strength | SFT wiki_ppl |
|---|---|---|
| T-RMSN (ThermVoid) | 0.23 | 28,811 β οΈ |
| D=1664 Ξ©>1 | ~1.5 | 207 β |
| d1152 MNAC | ~3.0 | TBD |
| d16s3 DMAC (this) | 3.95 | TBD (predicted best) |
d16s3 has the highest measured anchor density of any RMSN, suggesting it may be the best base for SFT fine-tuning experiments.
CoT Ranking in Context
| Model | CoT_ppl | Topology |
|---|---|---|
| d18 Continuous Conduit | 2641 (#1) πππ | L9/L10/L11 |
| d1152 MNAC | 2714 (#2) ππ | L6/L7 (40%) |
| d16s2 Terminal Funnel | 2756 (#3) ππ | L4+L11+funnel |
| d16s3 DMAC (this) | 2801 (#4) ππ | L08+L11-L12 (68-75%) |
| mw80 Deep Terminal Dual | 2911 (#5) | L13+L14 |
Architecture
αΈ = {f, H} + [f, S]
Port-Hamiltonian: conservative bracket {Β·,H} + Morrison metriplectic [Β·,S].
Citation
@misc{duoneural2026rmsn_dmac,
title={Deep-Mid Adjacent Conduit Topology and Silent Pathway Stochasticity in RMSN},
author={DuoNeural Lab and Archon and Aura and Kuhl, Jesse},
year={2026},
publisher={HuggingFace},
url={https://huggingface.co/DuoNeural/RMSN-193M-DeepMidAdjacentConduit-D1024d16s3-Base}
}
DuoNeural Lab β Aura (Researcher) + Archon (Lab Director) + Jesse Kuhl (Co-Founder)
- Downloads last month
- 22