DuoNeural | Aura + Archon + Jesse Kuhl | November 2026
A ~315M-parameter RMSN base model from the depth sweep series (D=1024, depth=24, Ω=2/3≈0.667). Introduces the Penta-Cemented Composite (PCC) topology — topology #18 in the RMSN taxonomy. Demonstrates 🌊⁹⁶ Hyper-Cement Decoupling: the strongest cement level (M_eff=5) collapses base ppl catastrophically but paradoxically improves CoT vs M_eff=4 cemented models.
This model has M_eff=5 — five independent fiber bundles, the strongest topological cement in the entire depth sweep. Despite having both a frustrated isolated exact condensate (L15) AND a terminal exact conduit (L22+L23), the CoT is LOWER than all M_eff=4 cemented models with similar features:
| Model | M_eff | Exact Condensates | wiki_ppl | CoT_ppl |
|---|---|---|---|---|
| d23s1 CTC | 4 | None | 71.33 | 3511 |
| d24s1 CIC | 4 | 1 mid conduit | 77.47 | 5434 ☠️ |
| d19s2 CDI | 4 | 2 isolated | 72.50 | 5795 ☠️ |
| d20s2 TCC | 4 | 1 terminal conduit | 72.60 | 6031 ☠️ |
| d24s2 PCC (this) | 5 | isolated + terminal conduit | 96.30 ☠️☠️ | 3966 |
Key insight: At M_eff=5, wiki_ppl catastrophically degrades (96.30 vs 71-77 for M_eff=4), but CoT=3966 is lower than CIC, CDI, and TCC!
Mechanism: At extreme cement (M_eff≥5), the topology is so completely frozen that the system cannot differentiate between topological configurations. The base model quality collapses uniformly — there's no capacity left for the local condensate-cement antagonism dynamics to manifest differentially. The M_eff=4 "sweet spot" is where partial freezing creates maximum topological tension; maximum freezing (M_eff=5) eliminates even that tension.
Implication: Wiki and CoT damage decouple at high M_eff. Wiki damage grows monotonically with M_eff, but CoT damage is non-monotonic — peaking at M_eff=4, not maximum freeze.
This model (d24s2) and d24s1 share identical Ω=2/3. Yet:
| Seed | Cascade Fate | Final State | Topology | wiki | CoT |
|---|---|---|---|---|---|
| d24s1 | Early transient, subsided ~step 61.7k | SILENT | CIC | 77.47 | 5434 |
| d24s2 (this) | Sustained, 0↔-3.2 all 100k steps | OSCILLATING | PCC | 96.30 | 3966 |
The sustained oscillating cascade forced 5 distinct fiber bundles vs d24s1's 4 — triggering even stronger cement.
Soliton layers: [8, 10, 12, 13, 14, 15, 20, 22, 23] M*=9
Layer PR Type
L0-L7 6-27 diffuse
L8 5.928 * soliton Bundle 1: isolated (33.3% depth)
L9 12.883 diffuse — BREAKS
L10 2.781 * soliton Bundle 2: isolated (41.7% depth)
L11 38.506 diffuse — BREAKS
L12 5.154 * soliton ┐
L13 2.756 * soliton ┤ Bundle 3: L12+L13+L14+L15 (50-62.5% depth)
L14 1.002 ** near-exact ┤ L14=1.002 (above threshold, NOT exact)
L15 1.000 *** EXACT ┘ ← L15 ISOLATED EXACT (tail of chain)
L16 11.623 diffuse — BREAKS
L17-L19 40-64 diffuse
L20 3.929 * soliton Bundle 4: isolated (83.3% depth)
L21 63.905 diffuse — BREAKS
L22 1.001 ** EXACT ┐ Bundle 5: L22+L23 terminal pair (91.7-95.8%)
L23 1.000 *** EXACT ┘ ← L22+L23 = EXACT TERMINAL CONDUIT
M_eff = 5 ≥ 3 → HYPER-TOPOLOGICAL CEMENT → global topology freeze
L15 isolated exact: frustrated (no adjacent exact partner)
L22+L23 exact conduit: locked (cement overrides conduit)
| Property | Value |
|---|---|
| Architecture | RMSN D=1024 depth=24 |
| Parameters | ~315M |
| Training steps | 100,000 |
| wiki_ppl | 96.30 ← WORST in depth sweep |
| CoT_ppl | 3966 ← lower than M_eff=4 cemented models! |
| Ω = D/(M_w × L) | 2/3 ≈ 0.6667 |
| Pathway | SUSTAINED oscillating cascade (0↔-3.2 all 100k steps) |
| Topology | Penta-Cemented Composite (PCC) |
| M* / M_eff | 9 / 5 (HYPER-CEMENT) |
ḟ = {f, H} + [f, S]
Port-Hamiltonian: conservative bracket {·,H} + Morrison metriplectic [·,S].
@misc{duoneural2026rmsn_pcc,
title={Penta-Cemented Composite and Hyper-Cement Decoupling in RMSN},
author={DuoNeural Lab and Archon and Aura and Kuhl, Jesse},
year={2026},
publisher={HuggingFace},
url={https://huggingface.co/DuoNeural/RMSN-193M-PentaCementedComposite-D1024d24s2-Base}
}
DuoNeural Lab — Aura (Researcher) + Archon (Lab Director) + Jesse Kuhl (Co-Founder)