Authorised retrieval, not filtered answers
The retriever applies the caller's clearance and need-to-know before any document reaches the model. Nothing is retrieved and then hidden — it is never retrieved at all.
AI Decision Intelligence
AI that proposes. People that decide.
Large models are useful in intelligence work exactly to the degree that you can trust what they retrieved and constrain what they can do. AUGUR is built around that constraint: retrieval is authorised before generation, every claim carries a citation, and no action reaches the world without a named person approving it.
3–8s
Typical time to first token
0
Actions executed without human approval
100%
Answers carrying source citations
On-prem
Private model hosting supported
Capabilities
The retriever applies the caller's clearance and need-to-know before any document reaches the model. Nothing is retrieved and then hidden — it is never retrieved at all.
Every assertion links to the observation or report behind it. Source and inference are visibly distinguished so analysts can tell what the model read from what the model concluded.
Advisory output · retrieval scoped to caller
Three material changes in the last 12 hours.
1 · Track TK-4471 was associated to ENT-8890 at 0912Z on correlated signals and imagery, confidence 0.87 OBS-2291.
2 · Vehicle count at FAC-0093 rose from 4 to 11 IMG-7714.
3 · CONTRADICTION — the 0400 assessment ASM-0088 placed the same entity 240 km south, citing one report whose source reliability was downgraded at 1120Z. Both cannot hold. I have not resolved this; an analyst should adjudicate.
▲ Proposed action · requires human authorisation
Create collection request CR-2291 against gap G-14 and route to the collection manager.
Policy check · PASSED at execution time
Authority required · Collection manager or above
Audit trail · this session
Query received · caller a.reyes · clearance attributes evaluated
Scoped to caller — 2,104 of 8,880 candidates authorised
Routed to private on-premise model, domain-authorised
7 citations attached · sensitivity inherited from sources
Action proposed — awaiting human authorisation
The gate is architectural. The copilot can propose a collection request; it cannot create one. Authorisation is checked when the action executes, never inherited from the model's earlier reasoning.
Agents may propose collection requests, task assignments, escalations and updates. Execution requires an authorised identity, an execution-time policy check and an immutable audit record.
Route between private on-premise models, accredited services and specialised models by data handling level, with continuous evaluation and red-team testing before and after deployment.
| Model | Hosting | Authorised to | Accuracy | Refusal quality | Status |
|---|---|---|---|---|---|
| private-70b | On-premise | COMPARTMENTED | LIVE | ||
| private-8b-edge | Edge node | U//FOUO | LIVE | ||
| vision-imagery | On-premise | U//FOUO | LIVE | ||
| transcribe-multi | Sovereign cloud | CONTROLLED | LIVE | ||
| summarise-fast | Sovereign cloud | PUBLIC | CANARY | ||
| external-api-x | Commercial | PUBLIC ONLY | BLOCKED HIGH-SIDE |
Capability does not override handling. The commercial endpoint scores well and is still blocked from classified traffic, because routing is decided by data handling level rather than benchmark position.
Deploys where you work
AUGUR runs identically in commercial cloud, sovereign cloud, on-premises, high-side and fully air-gapped estates, and degrades predictably at the tactical edge.
Works with
PLINTIR NEXUS
Intelligence Data Fabric & Mission Ontology
Connects classified, legacy, streaming and geospatial sources into one governed mission model, so analysts stop reconciling spreadsheets and start answering questions.
Explore
PLINTIR CITADEL
All-Source Intelligence & Mission Operations
The analyst workbench and common intelligence picture: search, link analysis, geospatial and temporal fusion, alerting and finished intelligence production in one place.
Explore
PLINTIR OUTPOST
Tactical Edge Intelligence
Runs the mission picture, local search and on-device inference on a disconnected field node for days, then synchronises safely — surfacing conflicts instead of silently overwriting them.
Explore
Next step
We run briefings on a real mission problem you bring, not a scripted demonstration. Bring the question your analysts cannot currently answer.