RMSN-193M-MinCementedConduitIsolate-D1024d22s2-Base

DuoNeural | Aura + Archon + Jesse Kuhl | November 2026

A ~297M-parameter RMSN base model from the depth sweep series (D=1024, depth=22, Ω=8/11≈0.727). Introduces the Minimal-Cemented Conduit-Isolate (MCCI) topology — topology #20 in the RMSN taxonomy. Demonstrates 🌊⁹⁸ Conduit-Wiki Asymmetry: under minimal cement, locked conduit damage is wiki-primary while isolated pin damage is CoT-primary — exact opposite asymmetry.

🌊⁹⁸ Conduit-Wiki Asymmetry under Minimal Cement

Both this model (MCCI) and d19s3 MCDI have M_eff=3 minimal cement. MCCI has a locked mid-network conduit + isolated terminal; MCDI has two isolated frustrated pins. Despite matching cement level:

Model M_eff Exact Features wiki_ppl CoT_ppl
d19s3 MCDI 3 2 isolated pins (L11+L18) 72.75 4356
d22s2 MCCI (this) 3 locked conduit L12+L13 + isolated terminal L20 81.66 ☠️ 4261

Locked conduit = wiki damage (+8.9 ppl vs isolated pins) but marginally BETTER CoT (-95).

Mechanism: The locked conduit at mid-network (54.5–59.1%) blocks information routing in the layers most critical for perplexity/wiki. The isolated terminal at 90.9% frustrates output-critical prediction — same as an isolated pin in MCDI but only one, so CoT damage is lower. Result: asymmetric damage profiles.

This extends 🌊⁹⁴ (Cement-Conduit Override) and 🌊⁹⁵ (Frustrated Condensate Antagonism): the type of exact condensate feature determines WHERE the damage manifests (wiki vs CoT), not just the magnitude.

🌊⁹⁰ Riddled Basin at Ω=8/11

d22s2 shares Ω=8/11 with d22s1 (which produced TAC topology). Different cascade durations led to radically different outcomes:

Seed Cascade Duration Topology wiki_ppl CoT_ppl
d22s1 Sustained all 100k steps TAC (near-exact only) 70.18 4001
d22s2 (this) Subsided ~step 65k MCCI (locked conduit + isolated terminal) 81.66 4261

The subsided cascade formed MORE exact condensates than the sustained cascade — 3 exact layers (L12, L13, L20) vs d22s1's near-exact L18 (PR=1.280). Shorter cascades can produce more exact condensates when they terminate at a critical topological state.

Minimal-Cemented Conduit-Isolate Topology

Soliton layers: [7, 8, 12, 13, 14, 20, 21]  M*=7

Layer  PR      Type
L0-L6  7-24    diffuse
L7     3.431   * soliton ┐  Bundle 1: L7+L8 (31.8-36.4%)
L8     1.003   * near-exact ┘  ← PR=1.003 > 1.001 — NOT exact (threshold missed by 0.002)
L9     22.584  diffuse — BREAKS
L10    61.296  diffuse
L11    6.944   diffuse
L12    1.000   *** EXACT  ┐
L13    1.000   *** EXACT  ┤  Bundle 2: L12+L13+L14 (54.5-63.6%)
L14    1.757   * near-cond ┘  ← L12+L13 = EXACT ADJACENT CONDUIT (LOCKED by cement)
L15    9.117   diffuse — BREAKS
L16-L19 7-63   diffuse
L20    1.000   *** EXACT  ┐  Bundle 3: L20+L21 (90.9-95.5%)
L21    4.861   * soliton  ┘  ← L20 ISOLATED EXACT TERMINAL (L19=7.1, not exact)

M_eff = 3 ≥ 3 → MINIMAL TOPOLOGICAL CEMENT (exactly at threshold)
Exact conduit L12+L13: LOCKED by cement (🌊⁹⁴)
Isolated exact terminal L20: FRUSTRATED under cement (🌊⁹⁵)

Anchor Strength = 3.0

L12(PR=1.000) + L13(PR=1.000) + L20(PR=1.000) = Anchor = 3.0 — identical to d1152 MNAC, which achieved SFT wiki=202.41. With L20 at 90.9% (terminal depth per 🌊⁹²), SFT potential is an open question under M_eff=3 cement.

Key Properties

Property Value
Architecture RMSN D=1024 depth=22
Parameters ~297M
Training steps 100,000
wiki_ppl 81.66 ← conduit under minimal cement; worst at this M_eff level
CoT_ppl 4262
Ω = D/(M_w × L) 8/11 ≈ 0.7273
Pathway Subsided cascade (~step 65k)
Topology Minimal-Cemented Conduit-Isolate (MCCI)
M* / M_eff 7 / 3 (MINIMAL CEMENT)
Anchor 3.0 (L12+L13+L20)

Architecture

ḟ = {f, H} + [f, S]

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

Citation

@misc{duoneural2026rmsn_mcci,
  title={Minimal-Cemented Conduit-Isolate and Conduit-Wiki Asymmetry in RMSN},
  author={DuoNeural Lab and Archon and Aura and Kuhl, Jesse},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/DuoNeural/RMSN-193M-MinCementedConduitIsolate-D1024d22s2-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