Private deployment. Stronger proof.
Refinery
Run the same Kenshiki system in a private environment — without the public-model boundary.
Refinery is Kenshiki's private deployment tier. It keeps the full Kenshiki control plane but removes the public-model boundary. Refinery can run as a shared or dedicated instance in Kenshiki-managed AWS, inside a customer VPC or GovCloud account, or on premises when full air-gap is not required. In every mode, the answer path stays inside a private runtime: prompt compilation, retrieval, generation, claim evaluation, and output gating all happen under controlled infrastructure.
Without this: you can move AI into a private environment and still get answers no one can defend. Data residency solves where the model runs. It does not solve whether the output holds up.
Today
Your team moved AI into a private environment to satisfy data-residency, confidentiality, or procurement requirements. The model is no longer public, but the output is still just fluent prose that humans have to interpret and defend manually. When challenged, you can say where it ran. You still cannot show why a specific claim should be trusted.
With Refinery
The request now runs through Kenshiki inside a private deployment. The prompt is compiled, evidence is retrieved from governed sources, a private inference engine generates a proposal, and the Claim Ledger checks that proposal against evidence and local telemetry before assigning an output state.
How Refinery works
Refinery runs the same bounded-synthesis pipeline as Workshop, but moves generation into a private deployment. The prompt is compiled, governed evidence is retrieved, a private inference engine produces a proposal, and the Claim Ledger uses source checks plus local telemetry to decide what is allowed to leave.
Output states
What Refinery is
A private deployment of the full Kenshiki stack. It uses the same Prompt Compiler, retrieval, Claim Ledger, and output-state contract as Workshop, but generation happens on a private inference engine instead of a public endpoint.
- Private deployment tier for production workflows
- No public model API in the critical path
- Same bounded-synthesis contract as Workshop
The Kenshiki contract
Same contract. Private runtime.
Refinery runs the same Kura/Kadai contract as the rest of the platform. Kura defines what counts as real. Kadai is the answer contract the caller sees. The difference in Refinery is that the backing inference runtime is private: no public model API in the critical path, and no release without Claim Ledger evaluation.
- Same Kura/Kadai contract as Workshop
- Generation moves into a private runtime
- Claim Ledger evaluates both source support and local model telemetry
Who this is for
The Platform Team
deploying AI into private infrastructure while preserving data control, enterprise integration, and repeatable proof of what the system relied on.
The Reviewer
receives an answer that is already classified, traceable, and fit for use in production workflows — not raw model prose that still has to be defended by hand.
Go deeper
Claim Ledger
The verification engine inside every Refinery response. Breaks answers into claims, checks them against evidence, and records what held up.
Platform Architecture
See how Kenshiki runs the same contract across Workshop, Refinery, and Clean Room while changing where generation and proof occur.
Oracle Foundry
The governed source layer behind Kura — ingestion, provenance, chunking, and retrieval-ready evidence inside a private deployment.