SOPHIA XT
The Stack · Open core · MIT

The Compositional
SystemCompose, don't pretrain.

Most systems chase capability by retraining bigger models. This one assembles it at inference: a frozen commodity base for fluency, a trained router to pick the right skill, sealed NCN specialist units for that skill, and disk-backed MemK memory for everything you've said before — composed, not retrained.

Every number on this page is a measured proof-of-mechanism on our own harness — not a SOTA or leaderboard claim. We show the mechanism works; we don't claim it beats the field.

99.4%
Trained-router accuracy
100%
Recall@1 of old facts
16×
Capacity savings (IA³)
MIT
Open core

The flow

One query. Four moving parts. Composed at inference.

A turn arrives. The trained router decides which sealed NCN specialist owns it and what MemK memory to recall. Both condition a frozen base that was never retrained — and the answer falls out. Nothing here required a training run. Watch the token travel the path.
COMPOSITION_PATH.svglive
Query
user turn arrives
Trained router
picks the specialist
NCN specialist
sealed skill unit loads
MemK memory
recalls what matters
Frozen base
commodity fluency, untouched
Answer
composed, grounded reply

Frozen base

Fluency you inherit. You don't pay to re-learn it.

Diversity, grammar, world-knowledge, tone — every commodity base already has them. Retraining a base to add one skill throws that away and re-buys it at enormous cost. The compositional move is to leave the base frozen and add the skill around it. The economics are the whole point: you rent fluency from a commodity model and spend your budget only where it's differentiated.

Never retrained

The base's weights are read-only at inference. No gradient ever touches them in this stack.

Fluency for free

Grammar, coherence, and breadth come from the commodity model — the expensive part you don't have to fund.

Swappable

Because nothing is baked into the base, you can move the same specialists + memory onto a newer base later.

Trained router

Breaking the base-invariant routing ceiling.

Picking the right specialist is the accuracy bottleneck of any compositional system. A cosine-similarity router — the obvious approach — tops out around 86% on held-out routing. A small trained router head breaks that ceiling at 99.4%. The robust, augmented head holds 98.8% in production conditions (88% bare). And it can be BASELESS — a router with no base model attached, validated at the same 99.4% — so routing decisions don't depend on the base at all.
Cosine router
the obvious baseline — the ceiling
86.2%
Trained router
breaks the ceiling on held-out routing
99.4%
Robust head (prod)
augmented; holds under production noise
98.8%
Bare head
without augmentation
88%
BASELESS router
no base model — routing is independent
99.4%

NCN specialists

Sealed skill units. Not a LoRA. Not an MoE expert.

An NCN is a densely-trained, sealed and encrypted unit of skill — portable as a single .ncn file. It is not a low-rank adapter bolted to one base, and it is not an expert living inside a monolithic MoE. It carries its own skill and travels with it. The seal is lock-and-key: with the right key it loads; with the wrong key, nothing comes out.

Sealed & encrypted

The unit's contents are encrypted at rest. The key gates loading — wrong key, no skill.

Densely trained

Trained to carry a real skill at density, not a thin low-rank delta. Capacity where the data earns it.

Portable

One .ncn file moves between hosts and bases. The skill goes with the file, not the base.

Not a LoRA, not an MoE expert

A distinct artifact: self-contained, composable, and independently distributable.

MemK memory

Memory that scales by disk, not RAM.

MemK is disk-backed, encrypted, and real-time: capacity grows with storage, not memory — ~1TB ≈ ~1T tokens. In long-conversation tests it recalls old facts at 100% while spending roughly 20× fewer tokens than stuffing the full history back into context every turn. Memory accumulates on disk; retrieval pulls back only the handful of tokens that matter.
100%
recall of old facts (long-conversation test)
~20×
fewer tokens than history-stuffing
~1TB ≈ ~1T
storage-to-token scaling
Encrypted
at rest, real-time read/write

Data-gated capacity

Capacity is worth paying for — only when the data is rich.

More parameters help only when there's enough data to fill them. On thin data, a small rank-8 module wins; the high-capacity not-low-rank module overfits. As data gets rich, the lines cross and capacity overtakes. The measured gap walks −0.346 → −0.083 → +0.040 — the sign flips exactly at the crossover. And a tiny IA³ module captures ~99% of the skill at 16× fewer parameters. Spend capacity only where the data earns it.

gap (low-rank − high-capacity): −0.346 → −0.083 → +0.040 · sign flips at the crossover

16×
fewer params (IA³) for ~99% of the skill
rank-8
wins on thin data
+0.040
high-capacity overtakes once data is rich

Anchor-token harvesting

Dense training data, harvested from models that already exist.

You don't always need fresh human data to train a specialist. Anchor-token harvesting extracts dense, high-signal training data straight out of an existing published model — a methodology we've published. Specialists trained this way carry a measured 1.67× out-of-distribution generalization advantage over the baseline.

Published methodology

The harvesting + anchor-token masking method is written up and reproducible. Read the research.

1.67×

measured out-of-distribution generalization lift over baseline — a proof-of-mechanism result, not a leaderboard claim.

Open core

Open framework. Proprietary skill layer.

The framework that makes composition work — the memory, the router, the composer — is open source under MIT as sophia. The differentiated layer stays closed: the sealed specialist units, the production-trained router, and anchor-token harvesting. Open scaffolding; proprietary skill.
Open · MIT · sophia
  • Disk-backed MemK memory engine
  • Compositional router interface + composer
  • Reference quickstart and SDK
  • Runs on CPU; works behind any OpenAI-compatible model
Closed · SOPHIA XT
  • Sealed, encrypted NCN specialist units
  • Production-trained router (99.4% / 98.8% robust)
  • Anchor-token harvesting pipeline (1.67× OOD)
  • Densely-trained, distributable skill catalog

Compose it yourself.

The open core is MIT and runs on a laptop. The papers document the mechanism. The Sophia Memory page is the fastest way in.