RMSN-193M-MinCementedTripleMidIsolate-D1024d23s2-Base
DuoNeural | Aura + Archon + Jesse Kuhl | November 2026
A ~306M-parameter RMSN base model from the depth sweep series (D=1024, depth=23, Ξ©=16/23β0.6957). Introduces the Minimal-Cemented Triple Mid-Isolate (MTCI) topology β topology #21 in the RMSN taxonomy. Demonstrates πΒΉβ°β° Terminal Void Clustering: when all exact condensates cluster mid-network with a complete terminal void, wiki ppl clusters at ~81.6 regardless of mid-network arrangement β and significantly worse than a model with a truly-terminal exact condensate.
πΒΉβ°β° Terminal Void Clustering (New Law)
d23s2 has three isolated exact condensates (L9, L12, L14) all clustering in the mid-network region (39β61% depth). The terminal region (L16βL22, 69.6β95.7% depth) is completely void β all layers show PR=34β64, the highest diffusion in the depth sweep. Despite M_eff=3 (same as MCDI d19s3), the wiki retention is significantly worse:
| Model | M_eff | Terminal Exact? | wiki_ppl | CoT_ppl |
|---|---|---|---|---|
| d19s3 MCDI | 3 | L18 at 100% depth | 72.75 β | 4356 |
| d22s2 MCCI | 3 | L20 at 90.9% depth | 81.66 β οΈ | 4261 |
| d23s2 MTCI (this) | 3 | None (terminal void) | 81.65 β οΈ | 4634 |
Finding: Under M_eff=3 minimal cement, wiki ppl WITHOUT a truly-terminal exact condensate clusters at ~81.6, while a model with an exact condensate at ~100% depth achieves wiki=72.75. The terminal exact-at-output-depth advantage applies to base pretraining retention, not only SFT fine-tuning.
Per-Pin CoT Penalty: Linear and M_eff-Scaled
MTCI confirms the per-pin CoT penalty is linear at M_eff=3 and M_eff-scaled:
| Config | M_eff | Exact Pins | CoT_ppl | Per-pin penalty |
|---|---|---|---|---|
| MCDI d19s3 | 3 | 2 isolated | 4356 | ~278/pin |
| MTCI d23s2 (this) | 3 | 3 isolated | 4634 | ~278/pin β |
| CDI d19s2 | 4 | 2 isolated | 5795 | ~1142/pin |
The per-pin penalty at M_eff=3 is β278 (confirmed by two independent data points). At M_eff=4 it jumps to β1142 β a ~4Γ amplification per M_eff step. This extends πβΉβ· (Cement-Scaled Frustration): the penalty scales super-linearly with cement level, not just linearly.
πβΉβ° Riddled Basin: Same Ξ©=16/23, Opposite Fates
d23s2 and d23s1 share identical Ξ©=16/23:
| Seed | Lyapunov | Final State | Topology | M_eff | wiki | CoT |
|---|---|---|---|---|---|---|
| d23s1 | SILENT all training | Quiescent | CTC | 4 | 71.33 | 3511 |
| d23s2 (this) | Oscillating -0.07β-0.19 all training | Sustained band | MTCI | 3 | 81.65 | 4634 |
The oscillating cascade produced 3 mid-network exact condensates but left the terminal region completely void. The silent pathway produced a cemented topology with higher M_eff (4 vs 3) despite less cascade activity β another example of the complex relationship between cascade dynamics and soliton proliferation.
Minimal-Cemented Triple Mid-Isolate Topology
Soliton layers: [6, 7, 8, 9, 12, 14, 15] M*=7
Layer PR Type
L0-L5 8-29 diffuse
L6 4.915 * soliton β
L7 2.996 * soliton β€ Bundle 1: L6+L7+L8+L9 (26.1-39.1%)
L8 2.254 * soliton β€
L9 1.000 *** EXACT β β L9 ISOLATED EXACT (L10=24.0, breaks chain)
L10 24.002 diffuse β BREAKS
L11 30.696 diffuse
L12 1.000 *** EXACT Bundle 2: L12 ALONE (52.2%) β isolated
L13 8.743 diffuse β BREAKS
L14 1.000 *** EXACT β Bundle 3: L14+L15 (60.9-65.2%)
L15 3.630 * soliton β β L14 ISOLATED EXACT (L13=8.743, breaks chain)
L16 53.424 diffuse β
L17 61.400 diffuse β€
L18 53.912 diffuse β€ TERMINAL VOID β zero solitons
L19 60.861 diffuse β€ (all PR=34-64, highest diffusion in depth sweep)
L20 63.866 diffuse β€
L21 34.463 diffuse β€
L22 63.875 diffuse β
M_eff = 3 β₯ 3 β MINIMAL TOPOLOGICAL CEMENT (exactly at threshold)
All 3 exact condensates: mid-network only (39-61%)
Terminal region (70-96%): COMPLETE VOID
Key Properties
| Property | Value |
|---|---|
| Architecture | RMSN D=1024 depth=23 |
| Parameters | ~306M |
| Training steps | 100,000 |
| wiki_ppl | 81.65 β no terminal exact condensate; baseline ~81.6 for M_eff=3 terminal-void |
| CoT_ppl | 4634 β 3 isolated mid-network pins @ ~278/pin penalty |
| Ξ© = D/(M_w Γ L) | 16/23 β 0.6957 |
| Pathway | Oscillating band cascade (lyap -0.07β-0.19 sustained all 100k steps) |
| Topology | Minimal-Cemented Triple Mid-Isolate (MTCI) |
| M* / M_eff | 7 / 3 (MINIMAL CEMENT) |
| Anchor | 3.0 (L9+L12+L14) β all mid-network; weaker SFT expected vs terminal anchor |
Architecture
αΈ = {f, H} + [f, S]
Port-Hamiltonian: conservative bracket {Β·,H} + Morrison metriplectic [Β·,S].
Citation
@misc{duoneural2026rmsn_mtci,
title={Minimal-Cemented Triple Mid-Isolate and Terminal Void Clustering in RMSN},
author={DuoNeural Lab and Archon and Aura and Kuhl, Jesse},
year={2026},
publisher={HuggingFace},
url={https://huggingface.co/DuoNeural/RMSN-193M-MinCementedTripleMidIsolate-D1024d23s2-Base}
}
DuoNeural Lab β Aura (Researcher) + Archon (Lab Director) + Jesse Kuhl (Co-Founder)
- Downloads last month
- 25