Use case
Grounding sovereign and public-sector agents
Government, regulators, public broadcasters and defence programmes running agents on a model they control, inside a boundary they control.
The problem, and the approach
The problem
- The model is sovereign; the retrieval is not. An agent on a national model that grounds itself through a foreign search API sends every question abroad, and the question is the intent.
- Assurance needs a record. "The agent searched the web" is not an auditable statement; a reviewer needs what was asked, by whom, what evidence came back and what it was told was missing.
- Nothing in the stack may decide what is true. A public body cannot deploy a system that labels content; it can deploy one that records who reported what and what corroborates it.
The approach
- ground returns provenance-tagged, deduplicated evidence with a coverage receipt and no generated answer, so the judgement stays with the model you run.
- independent_origins counts origins and owners rather than hosts, so a wire story and its forty reprints are one voice and the receipt says so.
- The sovereign edition deploys the whole layer — index, coverage graph, receipts — inside your boundary, and the query record is held by you, in full, attributed to the calling principal.
- why_not answers for any URL whether it is present, was removed and why, or was never admitted, so a challenged absence has a typed answer rather than silence.
The parameters that matter
| Parameter | Value | Why |
|---|---|---|
min_independent_origins | 2 or 3 | A corroboration threshold written into policy and applied to every query. |
source_roles | official, primary | Prefer the originating institution over reporting about it. |
max_age | the review window | The receipt says whether the evidence falls inside it. |
why_not?url= | any challenged document | A substantive answer to "why was this not considered". |
curl -H "x-api-key: $UNLOB_API_KEY" \
"https://api.unlob.com/ground?objective=flood+defence+funding+announcement&max_age=7d&min_independent_origins=2"The query is the disclosure
Public-sector procurement has learned to ask where data is stored. It has been slower to ask where a question is processed. A search API call is processed wherever the provider runs, and the query string is the operator's intent in the operator's own words: the entity of interest, the time window, the framing. A stream of them over a month is a map of an organisation's attention. Keeping the model at home while sending the questions abroad protects the less sensitive half.
This is why the sovereign edition is a deployment rather than a contractual promise. A provider can undertake not to retain queries; it cannot undertake not to have seen them. Inside the boundary there is no provider to see them, because the index is there too.
A record that is deliberately not anonymised
On unlob cloud the query record is scoped to a tenant and kept in aggregate and fingerprinted form, under the privacy policy. In a sovereign installation the opposite is right: the query is recorded in full, attributed to the calling principal, and joined to the receipt it produced, so a reviewer can replay the exchange exactly. That record is the assurance artefact — and it is the customer's, on the customer's retention policy, leaving the boundary never.
State this to whoever signs off the deployment as both a capability and a responsibility. The installation makes you the data controller of a fully identifying log of what your agents asked. That is what an audit needs, and it is what a foreign provider already holds about you today at the moment of every call. The question was never whether the queries are recorded; it is who holds the record.
What the layer refuses to do
unlob does not decide what is true and does not label content. The receipt says how many independent origins support a claim, which source classes are missing, and where a claim first entered the record; it does not say whether the claim is correct. That boundary is what makes the layer deployable by a democratic government or a regulator at all — the judgement is visibly the model's and the people's, not the infrastructure's.
Where this is not the right tool
- The sovereign edition is delivered per installation and scoped with you; it is not a plan on the pricing page and it is not switched on in the console.
- The index is bounded on purpose and smaller than the well-funded competitors'. Inside a boundary the source policy is yours, but if raw coverage of the open web is the binding constraint, say so early.
- A receipt is a technical record, not a legal one. Whether it satisfies a particular obligation is a question for your assurance function.
Frequently asked questions
Does any query leave our boundary?
Not in a sovereign installation. On unlob cloud a query is sent to us and handled under the privacy policy. Inside your boundary the query is processed and recorded there, and nothing — query, evidence, or telemetry — goes out. Index snapshots come in on a schedule you set.
Do we need to use your model?
There is no unlob model. The evidence layer returns structured evidence and a receipt, never a generated answer, so it sits beneath a national foundation model, an open-weights model on your compute, or a commercial one.
Does unlob decide what is true?
No. It records what was reported, by whom, where a claim originated, which independent origins corroborate it and what is missing, and hands that to your model. The judgement stays with your model and your people.
Scope it with us
The sovereign edition is deployed inside your boundary and priced per installation. Tell us the jurisdiction and the deployment shape.