Four stages, and only one of them involves a model.

Reading and structuring your knowledge uses AI. Answering does not — by the time an agent asks, the thinking is already done and recorded.

  1. stage 1 of 4

    Read

    We read what your company already produces — email, chat, documents, records. Nothing to migrate, nothing new for anyone to maintain.

  2. stage 2 of 4

    Structure

    Every rule and exception becomes a typed node carrying its source quote, who decided it, what authority they hold, who may see it, and when it took effect.

  3. stage 3 of 4

    Compile

    That structure compiles into skills your agents run — a different surface per role, generated from the graph with no hand editing.

  4. stage 4 of 4

    Guard

    Permissions are enforced inside the query before anything is generated. Consequential steps start as drafts. Contradictions wait for a named human instead of overwriting your policy.

This is not search, and it is not a chatbot.

Search finds you a document. A chatbot gives you a plausible paragraph. Neither can tell an agent that invoices above AED 10,000 go to the Managing Partner, that a recurring utility bill under AED 2,500 is exempt, or that the person who decided the exemption had the authority to. Writguard holds that as structured, permissioned, executable logic — and refuses to answer where it has no standing.

Your brain is a folder of files you own.

The source of truth is markdown in a git repository — one file per fact, with its provenance in the front matter. You can open it, diff it, review a change the way you review code, and take the whole thing with you. The database we serve queries from is a disposable projection: delete it and it rebuilds from the files.

One file per fact
Readable without us, and readable in ten years.
Every change is a commit
Who changed what, when, and what it was before.
The database is disposable
Rebuilt from the record layer. Nothing lives only in a system you do not control.
Portable by construction
Not an export button we might one day remove.

Permission is not a filter applied afterwards.

The same question, asked by two people holding different keys. The second call authenticated perfectly well. It was simply never eligible to select the row — which is why there is nothing left to leak.

What is the partner payroll figure for April 2025?

Released key partners_only

amount_aed: 48250.00

1 source · brain://node/payroll-2025-04-partner/event/evt_9c1a44b2

Nothing released key finance_team

amount_aed:

0 source references · no amount in text, projection, provenance, or trace

Verified on live Postgres, session S7.1.

See what we measured → every number traces to a frozen artifact