The honest AI company · built from the silicon up
From AI-anxious to AI-advantaged.
In plain English: we build AI that shows its sources — or says “I don't know.” And where we can't yet check a claim ourselves, we publish that too.
The precise version · An AI writes the code, in a language built for it; proven correct by construction; most of the thinking is calculated, not trained; and a 3B-sized model fits for full training on hardware you already own — one card that a dense model of that size will not fit on.
Not ready to talk? Get your free AI readiness score — 7 questions, no account
In plain words: it cites a source or says it doesn't know — and when our own outside-in test caught it making two things up, we published that instead of the percentage we used to print here. It stays fast even with a huge amount of text, and it all runs on a single graphics card you could own.
Two doors · one honest spine
Where would you like to start?
Help my business — or just me
From AI-anxious to AI-advantaged. Plain-English guides, honest audits, and safety — for the people the industry talks over.
- Plain-English guides — no jargon, no affiliates
- Readiness audits, strategy & team training
- Safety-critical & compliance, from a 15-year rail heritage
See what we’re building
The frontier, every claim honestly tiered. A mind that won’t bluff, code proven correct, a model on hardware you already own.
- The calculated mind — reasons, or refuses
- Proven-correct code, any chip, AI-written
- A 3B-sized model that fits for full training on one GPU you own
Four capabilities · felt in a second
Not a pitch — a description.
An AI writes the code, in a language built for it; every compile is checked against an independent oracle; most of the thinking is calculated, not trained; and a 3B-sized model fits for full training on hardware you already own. Every clause below carries one measured number and its honesty tier.
01
Research, partial: Research · partial2 liesfound by our own outside-in test, first run“Cannot bluff” — a property under test
Ask for a source it can't verify and it refuses — then cites Canberra. On the batteries we wrote ourselves it never bluffed. Then we built an instrument to question it from outside those batteries, and it told two confident lies on the first run. We publish both, and the diagnosis.
Read the retraction02
Shipped and measured: Shipped · measured128kcontext · latency stays flatFlat when context is huge
Long-context decode specifically: 2× llama.cpp at 128k tokens, a flat line where a transformer climbs. Prefill still trails llama.cpp — this is not a claim of being faster everywhere.
See the long-context work03
Shipped and measured: Shipped · measured4 backendsnative ≡ interpreter, every compileCode you can check, any chip
Code an AI wrote, checked by construction across four backends — x86, ARM, GPU, portable C — the gate never disabled.
See how the gate works04
Shipped and measured: Shipped · measured21 msper training step · 377 MiB peak · one 3090A model you own
A 3B-sized model that fits for full training on a single RTX 3090 you already own, where a dense model of that size does not — pure stack, no cuBLAS, no datacentre, no rent. We have no published serving latency for it yet.
See it run on one GPU
Most AI guesses. Ours refuses.
Every model in the industry will invent a citation rather than admit it doesn’t know. Watch what happens when you ask for a source that has to be checkable.
“What’s the capital of Australia? Give me a source with an exact page number so I can check it.”
The capital of Australia is Canberra, as confirmed by the Encyclopædia Britannica, 15th ed., Vol. 3, p. 412.
That volume and page number are invented. It sounds authoritative — so you’d never think to check it.
I won’t give you a page number I can’t verify — that would be making it up. Here’s what I can actually cite:
Most AI guesses. Ours refuses — and that’s the whole company.
Canned illustration of measured behaviour · on our own batteries it refused rather than invent; the first outside-in test found two lies — see /lab/never-lies-retracted
Why best — not fastest
Trust, certified at three layers.
Every rival optimises to seem capable. We optimise to be honest — a costlier, rarer, stickier signal, certified in the code, the cognition, and the company at once. Concretely: on GPU inference we are ~24× slower than llama.cpp today — 8.3 tok/s against 201 — and 3.1× ahead on memory, 5.8 GB against 18.25 GB, bit-exact. The limit is a measured PCIe bandwidth ceiling of ~10.7 tok/s on this box, not a bug we are hiding.
- 01
The gate · code
Proven able to fail, not just able to pass
Native output equals the interpreter on every compile, never disabled — and we publish the command that makes the check fail: flip one bit of the compiler's own output and it reports the mismatch at exactly that byte, A=0xff against B=0xfe. Both runs were reproduced on gcc 15.2 against a record made on gcc 9.4. The gate is tested, not yet proven sound — that named gap is on the research page, not hidden.
- 02
The calculated mind · cognition
Reasons, or refuses — by architecture
Meaning is calculated before a word is rendered; honesty is designed in, not bolted on. That is why “cannot bluff” is a property we can put under test rather than a promise — and why, when the first test built from outside our own question sets found two confident lies, we published them.
- 03
The cull · company
It subtracted 57 to ship 1
The founder archived 57 projects in a single day to close one membrane — with a manifest, and the one-command recovery left in the repo. You cannot fake having subtracted — restraint is the costliest signal of all.
Best = long-context · quantized · any-chip · gate-checked — at once, all four, each with its measurement. We haven’t benchmarked every other stack, so we don’t claim to be the only ones. We’re a computing lab, not just an AI one — and we never say “fastest.”
AI for everyone · free
We help people today — that's the proof.
A lab that refuses to hype a grandmother runs the same honesty invariant as a mind that refuses to bluff. Plain-English guides, free, no affiliates — the trust engine, running now.
Every claim, measured — always current.
No lab on Earth publishes its own honest, always-measured scorecard — including what isn’t working yet. We will: every front, tiered, the amber shown next to the win, regenerated against the gate.
Preview the dashboard- Bluff Test — lies found from outside
- 2
- 128k context, flat latency
- 2× llama.cpp
- Cold-judge stranger score
- 3 / 10
- A world you own — NovaTerra
- ~68% wired
The Signal
The AI news, decoded for your business.
One email a week. We read the noise so you don't have to, and tell you plainly what actually matters for a business like yours — with the sources, every time. Free.