THOMAS
GARREN
Independent researcher and systems engineer
Founder-built core stack spanning production enterprise software, agent infrastructure, memory, routing and language-model research.
Built without a lab behind it
SOPHIA XT started outside a venture-backed lab and outside a corporate incubator, and it still runs that way. That is a constraint on what gets built and how fast, not a claim about anyone else. The work below is what came out of it, and every line of it opens to something you can check.
4 of the 9 public repositories are under Apache 2.0. LTMi-XT v0.1 is the retrieval format, with its reference implementation and benchmark corpora. Cassandra T1 is a masked-diffusion model shipping a verified epoch 5 FP16 checkpoint, the tokenizer and the architecture, at 28 layers and a hidden size of 2,048. lens-xt is a spec language for deterministic generation; at commit 01e5eed its suite runs 256 passed, 2 skipped, 1 failed, and the one failure is a stale assertion in the variant-registry test rather than a fault in the library. The anchor-token masking study behind Cassandra is published with its experiment log, its negative results and a reproduction recipe, deposited at DOI 10.5281/zenodo.20496783 under CC-BY-4.0.
7 systems are in daily use at companies named on work, each with what was delivered and what has not been verified stated on its own page. The internal tooling, with the measurement that settled each decision and the parts that remain unproven, is at built. Every figure on this page is carried from one internal source that records what measured it and when it was last checked, and each one is printed below with that method and the page you can check it on. Where a figure had no method behind it, it was taken off the page rather than rounded up.
The Product Portfolio
Each line opens a page on this site, a public repository, or a product with a URL. Client systems are at work.
24 papers, specifications and reports
14 of them download as a PDF, each size measured on disk. Every one is listed with its category, its licence where it has one, and the page it is presented on, at papers and downloads.
Cassandra T1
The masked-diffusion model released under Apache 2.0. The verified configuration is 28 transformer layers at a hidden size of 2,048, a 32,768-token vocabulary, grouped-query attention at 16 query and 4 KV heads, and a shipped epoch 5 FP16 checkpoint. The specification with its artifact hashes is served at /api/models.json. Output quality still needs more training and structured evaluation before production use, and the repository says so.
Anchor-token masking
Three masking objectives compared on identical data. Anchor-token masking gives a 1.67x held-out advantage over random-position masking with a 95% bootstrap interval of [1.51, 1.85]. The in-domain advantage is 1.26x at [1.18, 1.34], and the intervals do not overlap. Four mechanism probes end in a V/O ablation that locates the effect. The paper, the technical report and the reproduction recipe are all published, deposited under DOI 10.5281/zenodo.20496783, CC-BY-4.0.
Read the paperResearch Areas
What was measured, and by what
Each of these comes off a run, a test suite, a release manifest or a file on disk. Where a figure has no method behind it, it is not here.