Compositional Language Modelsthe first AI architecture that grows by composition, not retraining.
SOPHIA XT defined a new architecture category: the Compositional Language Model (CLM). A small shared base coordinates with proprietary NCN Module Specialists(.ncn files) through a capsule lattice — adding capability by composition, not by retraining the base. The same mechanism runs live on two different transformer architectures with zero code changes. The wrapper is real, the format is proprietary, and the result is measured.
What Is It
A new architecture, named and defined.
Categories belong to whoever names and ships them. SOPHIA XT named CLM the same way Vaswani 2017 named the Transformer and Jacobs 1991 named MoE — by defining the framing, shipping the proof, and putting the live demo on the public internet.
torch.load, safetensors, json, zip, tar — license-gated, license-checked.Why It's New
It's not a fine-tuning trick. It's an architecture.
- • Adding a new domain → retrain the whole model
- • Cost per domain: $10K–$1M, days to weeks
- • Routing is opaque (whatever the embeddings happen to do)
- • Cannot deploy at the edge (GB to TB of weights)
- • Adding a new domain → train one NCN Module (.ncn)
- • Cost per Specialist: ~$0.50, ~30 min on a consumer GPU
- • Routing is interpretable — cosine scores visible per query
- • Edge-deployable: 720 MB base + ~16 MB per Specialist
The category change is structural. CLM isn't a way to fine-tune — it's a way to compose. The mechanism is cosine retrieval over learned capsule embeddings, not a learned router; that's why it transfers across base architectures without code changes. The economic unit is the NCN Module Specialist, sold (or licensed) as a .ncn file. The operating model is compositional maintenance: surgical Specialist updates, not full retrains. We name it CLM because the right way to describe what we're doing isn't "another LLM." It's a new category.
How We Got Here
LTMi → LCMi → LCM → LCLM
The CLM didn't appear overnight. It's the result of a multi-year R&D arc starting from long-term memory infrastructure, through compositional mechanism design, into the operating model that shipped, into the full LCLM architecture currently in research.
Try It Yourself
Same wrapper. Two different base architectures.
Pick a query. The left column shows Mnema 360M live — the CLM mechanism running in production at sophiaxt.com/mnema. The right column shows Llama 3.2 1B wrapped with the same mechanism — actual measured routings from our 2026-06-05 cross-base benchmark. Both base architectures route through the same capsule lattice.
d_model=2048 capsules; arch-general capsule cosine is 0.847 (borderline-discriminative). Both bases agree on 8/10 queries — same routing mechanism, two different base architectures, zero code changes.The Proof
Every claim, measured.
We don't ship claims that aren't measured. Below: every load-bearing assertion we make about the CLM architecture, the measurement that established it, and the artifact (live page or paper) where you can verify it yourself.
| Claim | Method | Result |
|---|---|---|
| Mnema 360M routes correctly at N=2 | 10-prompt canonical benchmark (lclm_benchmark_extended_v1) | 10/10 correct |
| Cross-base wrapper: Llama 3.2 1B at N=2 with zero code changes | Same CLM mechanism on a different transformer — same benchmark | 8/10 correct (80%) |
| Edge-deployable (CPU, no GPU) | Live demo on 4-core CPU VPS, bf16 + KV cache | ~7 tok/s, 720 MB base + ~16 MB per Specialist |
| Per-Specialist training is capital-efficient | Composer training pipeline (sophiaxt-train CLI) | ~$0.50 in compute, ~30 min on a consumer GPU per Specialist |
| Anchor-Token Masking improves OOD generalization 1.67× | Peer-published technical report, Zenodo DOI | 1.67× OOD advantage · bootstrap CI [1.51, 1.85] |
| Proprietary .ncn file format resists standard ML tooling | Tested vs torch.load / safetensors / json / zip / tar | All 5 readers fail — opaque to commodity ML pipelines |
| N=6 Specialist routing via LDA discriminative subspace | Fisher LDA projection from 7-class labeled corpus → 6-d subspace, cosine retrieval on projected embeddings (no model retraining, pure math) | 25/30 = 83.3% accuracy · MI = 2.088 bits (vs RAW 1.633, NCN v3 0.279, NCN v2 0.187, NCN v1 0.000) |
Why We're First
Three things nobody else has done together.
.ncn file format..ncn files — opaque to torch.load, safetensors, json, zip, and tar. License-gated on load, license-checked at runtime. No existing adapter-routing system has a comparable enterprise-ready artifact format.Each of these is achievable individually. The combination — a named category with cross-base validation and a proprietary deployment format — is what makes the position defensible.
N=6 Breakthrough · 2026-06-06
We broke N=6 — via classical mathematics.
The outer composition layer (capsule lattice routing) was previously validated at N=2 Specialists. Scaling routing past 2 simultaneously-discriminated Specialists on similar-domain corpora was the open question. We tested four router-based approaches and all hit a saturation problem. Then we tested Fisher Linear Discriminant Analysis (LDA) — a 1936 mathematical technique applied to the existing capsule embeddings — and the saturation broke.
| Approach | Idea | Result (MI = mutual info, bits) |
|---|---|---|
| ρ re-extraction | Tune the capsule-mass parameter | 53% → 57% (marginal) |
| V2 top-K weighted blend | Blend top-2 instead of pick top-1 | top-2 ≡ top-1 (no headroom) |
| NCN router v1 (single layer) | Learn routing at one mid-layer | Collapse: all → node 7, MI = 0.000 |
| NCN router v2 (8 layers, unsupervised) | Multi-layer + anti-collapse losses | No collapse but per-prompt fixed (MI = 0.187) |
| NCN router v3 (8 layers + contrastive) | Add InfoNCE contrastive loss on 7-class supervised labels | Improved but plateaus (MI = 0.279) |
| RAW 960-d cosine (baseline) | Vanilla cosine retrieval on existing capsules | 17/30 = 56.7% accuracy · MI = 1.633 |
| LDA 6-d projection (BREAKTHROUGH) | Fisher Linear Discriminant — project to 6-d subspace where between-class scatter ≫ within-class scatter | 25/30 = 83.3% accuracy · MI = 2.088 bits · 0.744 of max |
The discrimination signal was already in the embeddings. The capsule centroids in 960-d hidden-state space contained the information needed to route 6 Specialists — we just needed to look at the embeddings in the right subspace. LDA projects to the K-1 = 6 directions where class separation is mathematically maximal. The Mnema base model never needed retraining; we needed the right linear projection.
The four router-based approaches failed because the routing architecture was trying to learnwhat was already present in the geometry. LDA finds that geometry directly. The next research direction — Stochastic-Evolutionary Plinko V6.5 — layers adaptive stochastic exploration on top of LDA's static optimal subspace, combining mathematically-optimal discrimination with entropy-guided online evolution.