Governed AI | The StreamBuild guide

If you are deciding

Nothing AI-drafted reaches a funder, an applicant or the public until one of your staff has put their name to each thing it guessed.

If you use it daily

The AI does the first draft. You see exactly what it read and what it assumed, and you settle those assumptions in a few clicks.

If you are doing diligence

One gateway for every model call; a manifest, extracted assumptions and a named owner per assumption on every artefact; the gate enforced server-side.

One gateway

Every AI call in the product goes through a single gateway. No screen, workflow or background job calls a model directly. That gateway is where the three things below are attached to the output, so they are present on every artefact, not only on the ones someone remembered to check.

What every AI-drafted artefact carries

1
A context manifest
The exact records the model was given, shown as chips: “Transcript”, “Engagement”, “Claim status”. If it was not on the manifest, the model did not see it.
2
Extracted assumptions
Each thing the model had to guess, written as a plain sentence a program manager can agree or disagree with.
3
A named owner per assumption
The person who accepted or corrected it, and when. An assumption nobody has accepted stays visibly unaccepted.

Capture review: one owned, two pending, frame 1 of 2 — FIG. 07 A meeting summary that exists but is not yet anybody’s, shown across two frames. The manifest records that it was given the transcript, the engagement and the claim status. One assumption has been owned and two are still pending, so filing is refused. The name of the person who accepted it and the time they did are stored, but this screen does not yet display them; the deliverable review below does. *The whole argument in one frame.*

Capture review: one owned, two pending, frame 2 of 2 — FIG. 07 A meeting summary that exists but is not yet anybody’s, shown across two frames. The manifest records that it was given the transcript, the engagement and the claim status. One assumption has been owned and two are still pending, so filing is refused. The name of the person who accepted it and the time they did are stored, but this screen does not yet display them; the deliverable review below does. *The whole argument in one frame.*

The gate

Output cannot be filed, sent or published while any assumption on it is unowned. That rule is checked by the server when the file, send or publish is requested, rather than by hiding a button. A screen that forgot to disable the button would still be refused.

Owning an assumption is an individual act. There is no “accept all”. Correcting one is a first-class choice: the person writes what is actually true, the draft updates, and the model’s original wording is kept in the record alongside the correction.

Deliverable review: filing blocked — FIG. 08 A generated report on an engagement. Filing is refused with the reason stated: two assumptions still need an owner. *The gate is real.*

Model tiers, and the refusal

Models are arranged in tiers, with automatic fallback when a provider has a transient failure. An agency can scope a model to a particular data residency. A model scoped that way will refuse to fall back to the default, US-hosted model: the request fails, and the person is told to configure a credential for a model in the right region.

We would rather the draft fail than move your records across a border without you deciding to.

AI settings: model tiers and residency — FIG. 09 The agency’s AI settings. Each tier names the model it uses and can be pointed at another one. Every tier here sits on the platform default, which is the US-hosted provider described above. *Model choice is the agency’s, per tier.*

What this does not do

The AI does not decide anything. It drafts; people own.
There is no bulk acceptance of assumptions, and no keyboard shortcut that owns one.
Residency is not automatic. The default model tier is US-hosted. Scoping a model to another region today means supplying your own endpoint. There is no first-party Canadian option yet. We are building that provider adapter, targeted for October 2026.
The manifest shows which records the model was given. It does not show the model provider’s own processing.