Technology
Verification is
the whole moat.
Anyone can generate code. The hard part is deciding, mechanically and without a human in the loop, whether what came out is actually right.
The model is ours, or it isn't a product
GeniusMoment will run on our own model, on our own infrastructure, with no third-party API in the path of a customer request. That is an architectural commitment rather than a preference we trade away under deadline: a product whose core capability is rented from a competitor is that competitor's feature. They price it and can deprecate it. Shipping one while claiming a thesis about hallucination would be selling a wrapper.
The client is already built that way. It speaks plain HTTP to an endpoint we operate,
with no vendor SDK anywhere in it. The wire format is
/v1/chat/completions because every open serving stack speaks it, including
vLLM, SGLang, llama.cpp, TGI, and Ollama. That is an interchange format, not a dependency. Point
it at our server and no third party sees the request.
Status, plainly: the serving stack is not up yet. The loop runs today against an open model on hardware we control, which is enough to build and measure it and keeps every request off a vendor's API. It is a workstation, not a service. Standing up a real GPU endpoint is the single gate in front of everything below, and it is what we are raising cloud capacity to do.
The loop is also a data generator
This is the part that compounds. Every round gm runs produces a (spec, code, verification outcome) triple, and the failed→repaired pairs are precisely the supervision a coding model wants and almost nobody has. Public corpora are full of finished code. They contain very little of the thing that actually matters: a wrong attempt, the exact reason it was wrong, and the fix.
So the product generates its own training set as a side effect of being used, and the training set makes the product better at the one turn that differentiates it, which is the repair turn.
The corpus today
- 510verified training records, every one of them executed
- 200held-out evaluation specs the training set never sees
- 186seed specifications across the request distribution
- 13phrasing shapes for the same underlying request
Nothing enters the corpus on a judgement call alone. A record is admitted only after its code has been executed and its tests have passed, in a sandbox with no network and no install step. Evaluation runs against a held-out set that is checked for overlap against every training prompt, so a score cannot be inflated by a spec the model has already seen.
The distribution is deliberate rather than scraped: alongside ordinary requests it carries a catalogue of repair-failure classes. It covers the twelve distinct ways a second attempt goes wrong, plus explicit negative cases such as requests that should be refused rather than answered.
What we run in the cloud
The workload is GPU-shaped from end to end, and it is the reason this is an infrastructure company rather than an application on top of somebody's API. This is what we are standing up, in the order we need it.
| Workload | What it needs |
|---|---|
| Inference serving | A coding model behind our own /v1/chat/completions endpoint, sized so a
multi-round repair loop stays interactive. |
| Fine-tuning | Supervised runs over the verified corpus, re-run as the corpus grows with real repair traces. |
| Verification fleet | Sandboxed, network-isolated execution of generated test suites. This is the highest-volume operation in the system, and it is pure CPU. |
| Evaluation | Held-out benchmark passes measuring pass rate and rounds-to-green per candidate model, which is a better signal than any public leaderboard. |
| Artifact storage | Durable object storage for run records, corpus snapshots and model weights, with integrity checks on every copy. |
Where we are honest about the limits
One model writing both the code and its tests can agree with itself. A test suite authored by the same context that authored the implementation shares its blind spots. Adversarial verification uses a second pass with a fresh context that sees only the spec and the code, then tries to break it. That is how the gap gets closed, and it is the next thing we are building.
We would rather say that on the technology page than let you find it yourself.
And what we refuse to build
Every capability points at something that does not exist yet. Not facts, not dates, not events, not citations, not biographies. Retrieval is a different product with a different failure mode, and it is the one place where inventing an answer is simply wrong. The boundary, in full →
Talk to us about the hard part
If you work on verification, sandboxed execution or code-model evaluation, we would rather hear from you than from another recruiter.
Get in touch