What is
SOPHIA XT?A Self-Organizing Extended Technology system.
SOPHIA XT is an arbiter compositional language system. Instead of pretraining one giant model to do everything, it composes a system from parts — a frozen base for fluency, a trained router, swappable per-domain specialists, and real-time long-term memory — and an internal arbiter sends each task to whichever engine is best for it.
The name
SOPHIA XT — wisdom, extended.
S-O-P-H-I-A. From the Greek σοφία, "wisdom." The emphasis on judgment — knowing which tool to use — is the whole point of the system.
Extended Technology. Capability is extended onto the existing model ecosystem — composed and arbitrated, not pretrained anew.
- Not Hanson Robotics' Sophia the robot .
- Not
sofiaxt.com— note the "ph" in SO PH IA XT.
The core idea
An arbiter compositional language system.
Frozen base
A commodity model supplies fluency. Never retrained in this stack.
Trained router
A small learned head decides which specialist owns each query.
NCN specialists
Swappable, sealed per-domain skill units — one .ncn file each.
MemK memory
Real-time, compressed long-term memory backed by disk.
The unit of growth is a small adapter, not a model. A new specialty is a new sealed file dropped into the system — the base, the router, and the memory keep working around it. See the full Compositional System for the mechanism in depth.
Model arbitration
The ARBITER routes each task to the right engine.
The best model for a hard reasoning step is rarely the best model for cheap, high-volume extraction. Arbitration lets each step pick its own winner.
The orchestration layer is built and wired, but it currently runs in a gated, dry-run mode pending operator activation. The plumbing is in place; the valve is deliberately closed. It is not yet executing live production traffic across all vendors — treat multi-vendor arbitration as engineered and ready, not as something running at scale right now.
NCN routing
How specialists get assigned — per query, not per token.
.ncn container. Routing happens per query, across distinct specialist files — the router picks which sealed unit owns a request; it is not choosing tokens one at a time inside a single model. The results below are early and honest — this is an in-development system, and every number carries its caveat.| approach | result | the caveat |
|---|---|---|
| Training-free cosine routing (N=2) | 18 / 20 | 10/10 Mnema + 9/10 published capsule build + 8/10 Llama-3.2-1B wrapper, with no code changes |
| Above N=2 (similar-domain) | collapses | top-1 falls 100% → 46.7% (N=2→6); a learned N≥3 router collapsed to mutual information = 0 bits. Open problem — reported, not hidden |
| Offline separability | ~82% to N=48 | the same signal separates ~82% of up to 48 targets across four frozen bases (Llama-3.2-1B / -3B, Qwen2.5-1.5B, SmolLM2-1.7B); the gap to deployment is the crux |
| Anchor-token masking (published) | 1.67× OOD | 95% CI [1.51×, 1.85×], n=48 · Zenodo DOI 10.5281/zenodo.20496783 |
The measured routing win today is training-free: cosine retrieval over capsule embeddings scores 18/20 at N=2 (10/10 Mnema, 9/10 on a published capsule build, 8/10 wrapping Llama-3.2-1B unchanged). Above two specialists on similar-domain capsules it collapses (100% → 46.7% by N=6), and a learned N≥3 router collapsed to zero mutual information — so the honest open problem is robust N≥3 routing. Offline, the same signal still separates ~82% of up to 48 targets across four frozen base models, which is why the gap to deployment is the crux, not the ceiling. Proof of mechanism — not a leaderboard claim.
The moat
Internal guidance, internal corpuses.
.ncn distribution format. The system's internal guidance is backed by internal, closed-source corpuses.Routing, memory & orchestration
The composition layer that decides which specialist, which memory, and which engine handles each task.
Curated, distilled corpuses
Domain data refined into dense, high-signal training material — internal and closed-source.
Encrypted .ncn distribution
Specialists ship as sealed, encrypted units. The format itself is part of the moat.
Setting expectations
What SOPHIA XT is not.
- A new frontier foundation model chasing raw benchmark scores.
- A finished, at-scale product — it is early-stage and pre-product.
- A drop-in replacement for the base models it composes.
- A composition layer on top of the whole transformer ecosystem.
- Focused on efficient, grounded, composable specialization.
- An honest, in-development system — proof of mechanism first.
Go deeper.
The definition is the surface. Each piece of the system has its own page.