RefineryDefine the purpose
Name the AI or workflow, the decision it makes, the approved source and fields, and what must remain excluded.
Four simple steps
From operational data to trusted AI context.
Define what an AI or workflow needs. Refinery reads the current source, checks whether the data is fit for that purpose, and returns only what may safely be used.
Read the source now
Refinery retrieves current PostgreSQL facts for the exact approved scope. Previous answers and cached assumptions are not enough.
Check what may pass
Freshness, completeness, identity, evidence, classification, and policy are evaluated for this exact purpose.
Return the decision
Trusted data is returned with evidence. Limited data carries its condition. Unsafe or uncertain data is blocked with a reason and next step.
Example: a support agent requests a customer's billing status. Refinery checks the approved purpose, current customer identity, data freshness, and field policy before the value reaches the agent.
Three possible gateway decisions
Trusted, limited, or blocked.
The objective is not to maximize data access. It is to return only what the declared AI or workflow may safely use and make the reason visible when it cannot.
Trusted
The data is current, in scope, policy-allowed, and supported by the evidence required for this exact purpose.
Limited
The data may be used only with a clear limitation, redaction, expiry, or required human oversight.
Blocked
The data is stale, conflicted, disallowed, or insufficiently proven. No raw fallback value is silently returned.
When something is wrong
A safe repair is a proof chain, not an AI answer.
Only eligible low-risk issues may enter the repair path. Identity conflicts, destructive changes, ambiguous financial facts, unsupported connectors, and missing authority stay blocked.
Detect
Identify the exact invalid or missing value using current source evidence and an agreed rule.
Propose
Prepare one bounded correction with its reason, subject, expected current value, and supporting evidence.
Approve
Require the applicable policy or a named human decision. Ordinary chat and model confidence grant no authority.
Write
Attempt the change only through the certified governed connector path for the exact approved subject.
Read back
Read the authoritative source again. A successful API response alone does not prove that the data changed correctly.
Receipt
Keep durable proof connecting the subject, authority, attempt, fresh readback, timing, and outcome.
What AI may help with
- Understanding the declared business purpose
- Summarizing current evidence and uncertainty
- Drafting a bounded repair recommendation
- Prioritizing the exceptions that matter most
What AI is never allowed to become
- Independent write or approval authority
- A replacement for fresh source readback
- Proof that a connector action succeeded
- A reason to hide identity conflicts or ambiguity
Current product status: Refinery is in controlled, PostgreSQL-first design-partner conversations. The gateway and repair contracts are implemented locally; general availability, target-host operation, and customer outcomes remain separate proof gates.
Start read-only
Check whether your AI use case fits.
We will map the decision, approved source, current risk, and evidence needed before discussing any governed repair.