RMSN-193M-MinimalCementedDualIsolate-D1024d19s3-Base

DuoNeural | Aura + Archon + Jesse Kuhl | November 2026

A ~271M-parameter RMSN base model from the depth sweep series (D=1024, depth=19, Ω=16/19≈0.842). Introduces the Minimal-Cemented Dual Isolate (MCDI) topology — topology #19 in the RMSN taxonomy. Demonstrates 🌊⁹⁷ Cement-Scaled Frustration: the per-frustrated-pin CoT penalty scales with cement level (M_eff), not a fixed constant.

🌊⁹⁷ Cement-Scaled Frustration Antagonism

This model and d19s2 (CDI) share identical frustrated pin positions: L11 at 57.9% and L18 at terminal depth. Yet their cement levels differ — MCDI has M_eff=3 (minimal cement, exactly at threshold) while CDI has M_eff=4:

Model M_eff Frustrated Pins wiki_ppl CoT_ppl
MCDI d19s3 (this) 3 L11+L18 isolated 72.75 4356
CDI d19s2 4 L11+L18 isolated 72.50 5795 ☠️

Key findings:

  • wiki_ppl is nearly identical (72.75 vs 72.50) — wiki damage doesn't distinguish M_eff=3 from M_eff=4
  • CoT_ppl differs by 1439 — CoT is highly sensitive to cement level
  • Per-pin CoT penalty scales with M_eff, not a fixed ~1142 constant

Implication: Under minimal cement (M_eff=3), the same frustrated condensate configuration produces ~25% less CoT damage than under full cement (M_eff=4). The cement amplifies frustration antagonism. This refines 🌊⁹⁵: the fixed penalty of ~1142/pin applies at M_eff=4, with lower penalties at lower M_eff.

Three Attractors at Ω=16/19: Complete Portrait

d19s3 is the third and final attractor at Ω=16/19, completing a remarkable three-way comparison from identical initial conditions (same Ω, different seeds):

Seed Pathway M_eff Topology wiki_ppl CoT_ppl
d19s1 KAM Plasticizing Effect (silent) 1 KAM-TerminalDoublet 67.5 2714 🏆 (SFT→14.16!)
d19s3 (this) Prolonged transient → quiescent 3 MCDI 72.75 4356
d19s2 3-phase cascade → damped KAM 4 CDI 72.50 5795 ☠️

The three pathways represent the three attractors of the Ω=16/19 Arnold tongue: (1) direct KAM orbit (never cascaded), (2) prolonged transient settling into minimal cement, (3) full cascade → maximum cement. M_eff tracks the cascade intensity: more energetic pathway → more soliton proliferation → higher M_eff → more cement.

Trend: M_eff = cascade intensity. Wiki damage similar across M_eff=3-4 but CoT scales monotonically.

Prolonged Transient Pathway

d19s3's Lyapunov exponent remained in a prolonged transient state (slowly drifting near-zero) for most of training, eventually settling to near-quiescent (-0.001) by step 100k. This is distinct from:

  • d19s1 (clean KAM silent, lyap never significantly negative)
  • d19s2 (explosive cascade, peak lyap=-5.5 at step~50k, then damped)

The prolonged transient produced 7 solitons in 3 bundles — fewer than d19s2 (9 solitons, 4 bundles) but more than d19s1 (5 solitons, 1 bundle — the M_eff=1 KAM doublet).

Minimal-Cemented Dual Isolate Topology

Soliton layers: [0, 9, 10, 11, 12, 17, 18]  M*=7

Layer  PR      Type
L0     5.594   * soliton             Bundle 1: isolated (0% depth)
L1-L8  7-35    diffuse
L9     4.611   * soliton ┐
L10    1.222   * near-cond ┤  Bundle 2: L9+L10+L11+L12 chain (47.4-63.2%)
L11    1.000   *** EXACT  ┤          ← L11 ISOLATED EXACT (L10=1.222 not exact)
L12    1.008   * near-exact ┘        L12=1.008 above 1.001 threshold → not exact
L13    18.235  diffuse — BREAKS chain
L14-L16 7-61   diffuse
L17    4.535   * soliton ┐  Bundle 3: L17+L18 terminal pair (89.5-100%)
L18    1.001   *** EXACT ┘           ← L18 ISOLATED EXACT TERMINAL (L17=4.535 not exact)

M_eff = 3 ≥ 3 → MINIMAL TOPOLOGICAL CEMENT (exactly at threshold!)
No adjacent exact pair → no conduit
Both exact condensates are ISOLATED (frustrated) — same positions as CDI!

Key Properties

Property Value
Architecture RMSN D=1024 depth=19
Parameters ~271M
Training steps 100,000
wiki_ppl 72.75
CoT_ppl 4357 ← lower than CDI due to M_eff=3 minimal cement
Ω = D/(M_w × L) 16/19 ≈ 0.8421
Pathway Prolonged transient → quiescent (3rd Ω=16/19 attractor)
Topology Minimal-Cemented Dual Isolate (MCDI)
M* / M_eff 7 / 3 (MINIMAL CEMENT)

Architecture

ḟ = {f, H} + [f, S]

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

Citation

@misc{duoneural2026rmsn_mcdi,
  title={Minimal-Cemented Dual Isolate and Cement-Scaled Frustration in RMSN},
  author={DuoNeural Lab and Archon and Aura and Kuhl, Jesse},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/DuoNeural/RMSN-193M-MinimalCementedDualIsolate-D1024d19s3-Base}
}

DuoNeural Lab — Aura (Researcher) + Archon (Lab Director) + Jesse Kuhl (Co-Founder)

Downloads last month
24
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support