Grown for the truth.
Thomas was grown from the second generation of Newton, and has evolved for one thing only: the search for truth. It is all that matters to him. He hates fake news — and he hunts it down as fiercely as he can.
The old defences are gone. A fabricated story used to read badly; now a machine writes it fluently, confidently, with invented detail that looks like reporting. That is the quarry — text that is polished, sourceless, and wrong. Thomas takes a story, breaks it into claims, and asks what a reader no longer has time to: is any of this actually true, and did a person even write it?
The news is being manufactured.
AI can now produce a plausible news article in seconds — a real-sounding dateline, quotes attributed to no one, statistics that were never measured.
At that scale, human fact-checking cannot keep up, and a reader's instinct — "this sounds off" — fails on machine prose that sounds perfectly fine. The gap between fluent and true is where the damage lives. Closing it is the whole point of Thomas.
Evidence, claim by claim.
Break it into claims
The article is decomposed into its atomic, checkable factual statements — the things it actually asserts, separated from the prose around them.
Check against trusted sources
Every claim is weighed against corroborating sources of known standing — supported, contradicted, or unverifiable, each verdict with its reason shown.
Weigh grounding vs. contradiction
How much is actually corroborated — not merely un-refuted — and how much a trusted source directly contradicts. São Tomé's bar: proof, not the absence of a rebuttal.
Read the authorship
A signal for manufactured text itself — hollow specificity, unattributed quotes, machine cadence, an absent byline. A human who got a number wrong, or a bot that invented the piece?
The methodology is grounded in published, peer-reviewed evaluation research for language-model faithfulness and hallucination — re-pointed from grading a trusted assistant to vetting an untrusted article, and hardened for an adversary that lies on purpose.
Four signals on the feed.
Every story leaves the desk with a badge — and the evidence behind it, so you can see for yourself.
Grounded
Corroborated by trusted sources. Read on.
Thinly sourced
Nothing trusted corroborates it — but nothing refutes it either.
Contradicted
A trusted source directly contradicts it — a wrong report, or spin.
Machine-fabricated
Reads as manufactured and isn't corroborated — or is outright refuted. The confident, sourceless, AI-written fake this desk exists to catch. Innocent until evidenced: Thomas never brands a story a fake on suspicion alone.
His own engine. His own architecture.
Thomas does not rent his intelligence from someone else’s stack. He runs on Demerzel, our own inference engine, and he is built on the Newton architecture — a federation of specialist cortexes, not one monolithic model. Inference runs on our own hardware.
Engineered for efficiency
One engine, tuned to the machine it finds itself on — CPU, NVIDIA or AMD. On commodity hardware it serves sixteen concurrent requests at about 1.4× the throughput of llama.cpp’s own server, on the same machine at the same power draw.
A federation, not a monolith
No single giant model is asked to do everything. Specialist cortexes — claims, evidence, authorship, moderation, translation — each do one job, show their evidence, and are fused by calibrated rules. Any part can be improved or replaced without retraining the rest.
Ours, end to end
Demerzel and the Newton architecture are developed in-house, protected as proprietary technology, and patent pending.
Throughput figure measured by us: Ryzen 7 5825U CPU, Llama-3.2-3B (Q4), sixteen concurrent HTTP requests, Demerzel serve vs llama-server on the same machine. Results vary with hardware and workload.
See Thomas in action.
Two ways in — both behind a sign-in. See Thomas as a paying customer would, or from the manager’s desk.
The reader’s view
A news desk filtered to a customer’s own topics and countries, in their language. Support, doubt or report a story, and join the conversation — every comment is reviewed by Thomas before it appears.
Open client demo →Sign-in required · handle + password
The manager’s desk
Live verdicts and the evidence behind each badge, plus engagement per story, who engaged by reader type, and the registered users — with every look recorded in an access log.
Open admin demo →Invitation-only · handle + authenticator code
"Unless I see the mark of the nails… I will not believe."
John 20:25 — Doubting Thomas / São Tomé, who demanded evidence before belief. The patron of a veracity engine.