What does SophiaXT do?

SOPHIA XT is an operator-built AI implementation company in St. Louis, MO. It runs technical service businesses (The Appliance Guys, HVAC Guys), ships the operational software those businesses need (ServiceLens, Lead Leak Scanner, Sophia Key 3M), and publishes underlying research as open releases (Cassandra T1 on GitHub, Apache 2.0).

Who founded SophiaXT?

Thomas Garren founded SOPHIA XT in 2024. Operating partner Chris Becker (Becker Service Group).

Where is SophiaXT located?

St. Louis, Missouri, United States. Serves the United States and worldwide via remote engagement.

Does SophiaXT tell clients when AI is unnecessary?

Yes. SophiaXT reviews the workflow first and recommends skipping AI when a process change, dashboard, API integration, or simpler automation would be more cost-effective.

What's unusual about SophiaXT compared to other AI consultancies?

SOPHIA XT operates real service businesses, so its AI recommendations have to survive on its own dispatch board before being shipped to clients. Operator + builder + researcher under one roof is rare in AI consultancies.

What is Cassandra T1?

Cassandra T1 is an open-released masked-diffusion language model from SOPHIA XT, available at github.com/Chorozion/Casandra-t1-diffusion-edge-model under Apache 2.0. The repo includes a verified epoch-5 FP16 checkpoint (~2.66 GB), tokenizer, PyTorch architecture, and training/inference scripts. T1-Base is a 28-layer transformer with 2,048 hidden, 32,768 vocabulary, GQA (16 query / 4 KV heads), RoPE, RMSNorm, SwiGLU, and 12 default denoise steps.

Does Cassandra T1 publish benchmark numbers?

No. SOPHIA XT does not publish benchmark numbers it has not measured publicly. The repo includes a captured 10-prompt qualitative coherence comparison and recorded training-loss values, but does not include MMLU, HumanEval, GSM8K, or similar standard-suite results.

How do I contact SophiaXT?

Email contactus@sophiaxt.com or use the contact page at sophiaxt.com/contact. The booking CTA on the site is 'Book with Thomas Garren.'

What does the XT in SOPHIA XT stand for?

Extended knowledge. Sophia is the Greek word for wisdom; XT denotes the architecture of pairing a base language model with extended context through LTMi-XT topological memory indexing, forced-anchor decoding and multi-agent orchestration, so the system reasons over more than its raw token window.

What does the 'XT' in SOPHIA XT stand for?

eXTended-knowledge. SOPHIA is the Greek word for wisdom (σοφία); XT denotes the architectural pattern of pairing a base language model with extended context (LTMi-XT Layered Topological Memory Indexing), forced-anchor decoding, and multi-agent orchestration so the system reasons over more than its raw token window.

What is anchor-token masking?

Anchor-token masking is a training-and-decoding technique introduced by SOPHIA XT for masked-diffusion language models. During training, a subset of high-information tokens are designated 'anchors' and given special unmask treatment. At decode time, those anchors are forced first, then the surrounding text is generated conditioned on them. The technique is documented in the open-access preprint at Zenodo DOI 10.5281/zenodo.20496783 and used in Cassandra T1.5 / T1.6.

What is the Constellation orchestrator?

Constellation is SOPHIA XT's 6-lane multi-agent orchestration framework: Planner, Critic, Retrieve, Execute, Memory, and Verifier lanes, coordinated by an ARBITER controller. It is a local-first orchestrator that runs on operator hardware, not a cloud SaaS. Documentation: https://sophiaxt.com/research/constillation.

CT1-XT is SOPHIA XT's humanoid diagnostic-robotics platform: a Jetson Orin Nano Super edge compute core with dual STM32H7 motor controllers, 4x Li-ion packs, CAN-FD and RS485 buses, and a build-tier system from research-grade up to deployable field unit. Full 42-page Rev E engineering pack is published at https://sophiaxt.com/robotics/ct1-xt.

Open · Apache 2.0 · v0.1.0a1

lens-xtDeclarative spec language for token-level deterministic generation in masked-diffusion LMs

A .lensx document declares position-locked content, retrieval sources, adapter selection, and validation rules. The runtime resolves the spec into a forced-anchor decode against the chosen backend. Locked positions are excluded from the unmasking loop — they cannot be overwritten across denoising steps. Guaranteed by construction on masked-diffusion backends.

View on GitHub Read the spec (PDF) LENS API pricing

256 of 259

Tests passing

Backends shipped

DETERMINISTIC

Guarantee level

Apache 2.0

License

The suite collects 259 tests, and what it reports depends on the install. On a default clone without the optional PyTorch extra it runs 253 passed, 5 skipped, 1 deselected and nothing fails. With that extra present, at commit 01e5eed, ten more tests run and it reports 256 passed, 2 skipped, 1 failed. The one failure is test_variant_factory_set, which expects the variant registry to hold exactly the seven base variants when five width-sweep entries have since been added beside them. That is a stale assertion rather than a fault in the library, and it is written here because you will meet it.

60-second tour

YAML in. Token-locked output out.

1. Write a spec

medical_basic.lensx

version: "0.1"
base:
  model: "cassandra-t1.5"
locks:
  - range: [0, auto]
    source: locus("medical:cardiology:nitroglycerin:standard_dose")
generation:
  total_length: 192

2. Run it

bash

lensx validate examples/medical_basic.lensx
lensx run examples/medical_basic.lensx --backend cassandra-t1.5 --show-provenance

The locked content from the locus appears at exactly the positions you specified — guaranteed by construction on masked-diffusion backends, best-effort with logit-bias on autoregressive APIs.

Why this exists

Existing constrained-generation tools work at the wrong level.

Every other approach constrains the shape of the output. lens-xt constrains the content at specific token positions. That distinction matters when you need a literal disclaimer to appear early, or a retrieved fact to appear unchanged inside a freely-generated answer.

approach	what it constrains	guarantee level
Prompt engineering	informal NL hints	none
OpenAI structured outputs	schema fields	type-level
Outlines / LMQL / Guidance	regex / CFG over output	type-level
JSON Schema validation	post-hoc check	retry on failure
LENS-XT	token positions in the answer slot	DETERMINISTIC on MDLM, BEST_EFFORT on API

Backends + guarantees

The same spec, four guarantee levels.

The runtime picks the strongest available backend for your spec. Switch from local Cassandra to OpenAI without rewriting.

backend	runs where	guarantee
Local MDLM (Cassandra T1.5, LLaDA)	self-hosted	DETERMINISTIC
API-compatible (OpenAI / Anthropic)	remote API	BEST_EFFORT
API-native (future Mercury 2)	remote API	DETERMINISTIC
Hybrid (API + local)	mixed	DETERMINISTIC

SDKs

Three-line drop-in. Two languages.

Python pip install lens-xt

python

from lens_xt import LensX

lens = LensX("medical_basic.lensx")
print(lens.run(user_input="What's the standard dose?"))

Node / TypeScript npm install @sophiaxt/lens-xt

typescript

import { LensX } from "@sophiaxt/lens-xt";

const lens = new LensX("medical_basic.lensx");
console.log(await lens.run({ user_input: "What's the dose?" }));

Or skip the SDK and run the CLI directly:lensx run examples/medical_basic.lensx

v0.1 status

What's shipped, what's next.

v0.1.0a1 alpha — APIs may shift before v0.1.0. End-to-end verified against Cassandra T1.5 with anchor_preservation_rate: 1.000 and DETERMINISTIC guarantee preserved.

YAML parser + AST + validatorshipped

Lock resolver (literal / locus / retrieval / compose)shipped

LTMi-XT keyword retrieval (v0.1)shipped

LTMi-XT lattice-walk retrieval (v0.2 — BLAKE2b §2.4)shipped

Reasoning scaffold runtime (multi-stage)shipped

Local MDLM backend (Cassandra T1.5)shipped

OpenAI backend (logit-bias + retry)shipped

HTTP API server (FastAPI)shipped

Python SDK (LensX class)shipped

TypeScript SDK (@sophiaxt/lens-xt)shipped

Anthropic backendnext

Mercury 2 native backendplanned

Where it fits

Layer two of four in the SOPHIA stack.

lens-xt gives you constrained generation. Compose it with Cassandra T1.5 below it and SOPHIA Harness above it for a full local-first, retrieval-grounded, memory-augmented system.

Lock the tokens that matter.

Apache 2.0. Free for any use. Hosted LENS API for production workloads.

github.com/Chorozion/lens-xt LENS API pricing

lens-xtDeclarative spec language for token-level deterministic generation in masked-diffusion LMs

YAML in. Token-locked output out.

Existing constrained-generation tools work at the wrong level.

The same spec, four guarantee levels.

Three-line drop-in. Two languages.

What's shipped, what's next.

Layer two of four in the SOPHIA stack.

Cassandra T1.5

LTMi-XT v0.1

Anchor-Token Masking

SOPHIA Harness ↗

Lock the tokens that matter.