Don’t trust us. Audit us.

You would be asked to connect a system to your firm’s email and documents, by a company you had not heard of last week. That is a real risk and we are not going to argue you out of it. What follows is what the architecture actually does, what we do not yet have, and what you can verify yourself.

Accurate as of 30 July 2026. If something here stops being true we change this page, not the conversation.

The controls, each with the thing you can check.

Access is granted and revoked by you

Connections to your systems are provisioned by your own administrator inside your own tenant, using your own identity provider. Revocation happens in your console and takes effect immediately. We hold no standing credential that survives your revoking it.

Revoke in your admin console and watch the next call fail.

Permissions are enforced before generation, not after

Every retrieval carries the caller’s scope into the SQL query itself. Out-of-scope rows are never selected, so there is no filtered-out content sitting in a context window waiting to leak. On our benchmark this produced zero permission leaks against a conventional baseline’s one.

Ask for the eval artifact. It is version-pinned and we will send the file.

One isolated stack per customer

Each customer gets a separate database and a separate container stack. There is no shared table with a tenant column. A cross-customer query is not forbidden by policy — it has no path to execute.

Architecture walkthrough on a call, including the compose files.

It runs where you choose

Our infrastructure, a server in your jurisdiction, or hardware you own. For strict residency the whole pipeline can run on local models, at a cost to answer quality we will quantify for you rather than hide.

Named deployment target written into the agreement.

Your knowledge is files you own

The source of truth is markdown in a git repository — one file per fact, provenance in the front matter, every change a commit. You can clone it, diff it, and keep it if you leave. The serving database is a disposable projection rebuilt from those files.

Read the repository. That is the whole verification.

Every access is logged, and the log is yours

An append-only record of every call: who asked, with which key, at what scope, which sources were returned, and which were withheld. It is written to your record layer, not to a dashboard we control.

Open the log file. It is in the same repository.

Untrusted content is quarantined before a model sees it

Content arriving from low-authority sources is classified and rejected at a deterministic boundary ahead of any model call, so a prompt-injection attempt in an inbound document does not reach a context window at all.

We will show you the quarantine test in the suite.

Agents propose; your people approve

Consequential steps compile as drafts with a named approver and an approval packet. Nothing irreversible commits without a human, until you explicitly relax that for a specific skill.

Approval packets are in the record layer with the decision attached.

Your data trains nothing

Zero-retention terms with our model providers, and no training on your content by us or by them. This goes in the contract rather than on a page.

Read the clause before you sign anything.

What we do not have.

Paperwork.

Data processing agreement
Available on request. Reviewed by your counsel, not ours.
Model terms
Zero-retention, no-training, named providers.
Security brief
One page, plain language, written for a partner rather than a CISO.
Deletion
Say the word and the stack is destroyed. You keep the repository.