Access
Gated, and built for agents.
Our engines are not a catalog you browse and not a service you can buy on a card. Access is requested, reviewed by a person, and granted with a bounded call quota. That is the whole commercial model — pricing is set per contract, settled with us directly.
What is behind the gate is designed to be driven by software rather than clicked through. There is a visual studio when you want to look at something, but the supported surface is the API and the CLI.
01
Request access
From the dashboard. The form asks for your organization, role, country, and intended use. That is the screening record, not marketing — this layer can touch export-controlled technology.
02
We review it
A person reads every request. No account is approved by default, and there is no tier you can buy your way into. We aim to respond within two business days.
03
Approval issues a quota
An approved account carries a call allowance for a period — monthly or annual, set against your contract. There is no balance to top up and no way to overspend: at zero, calls are refused until the period rolls or we raise it.
04
Mint a key, point an agent at it
Keys are issued from the dashboard once access is approved, prefixed cog_sk_, shown once. Start at the capability manifest — the engine is built to be read by a program before it is called.
Why an agent can use this safely
The engine states what may be believed.
An OpenAPI spec tells a caller what can be called. That is not the question an autonomous caller has to answer before it acts on a number. Every function we expose carries, alongside its schema:
- Fidelity — what the ground truth behind it actually is: closed-form physics, a surrogate of an analytic oracle, a surrogate of an EM solve, or an exact diagonalisation.
- Validation — the measured held-out error, the reference it was measured against, and the split used.
- Envelope — the input box the answer is trusted inside.
- Refusal conditions — when it returns a refusal instead of a number, so a caller can predict one rather than discover it.
Refusals are structured and repairable: an out-of-envelope request names the field, the value it got, and the range it allows, so an agent can fix the call without parsing an English sentence. A prediction outside its trained region abstains and says so rather than extrapolating quietly. Every answer is stamped with which engine produced it.