Experiment

Living Evidence Graph

Maintain a continuously improving representation of the people, places, organisations, interventions, evidence and outcomes Defra knows about.

AI-enabledFoundationalNow -> Next
KNOWREMEMBERUSEEXCHANGELEARNANTICIPATE

01

Problem to address

Defra’s real-world objects are highly connected, but systems tend to store transactions separately. People need to understand relationships, conflicts and change, not just retrieve records. At present those relationships are often reconstructed by people.

02

What the evidence tells us

Regulators need one picture of a producer. Field officers need to see how farmer, holding, parcel and agreement relate. Case workers rarely see the whole picture. With millions of data points, people also need help spotting which claims deserve attention first.

03

The proposition to explore

The idea

Represent important people, places, organisations, interventions and evidence as a living network of relationships with provenance, time and confidence.

04

Hypothesis

If Defra represents important entities and evidence as an explicit, provenance-rich graph, the same connected knowledge can support operational decisions, anomaly detection, user-facing views and policy learning without every team rebuilding joins.

05

How it could work

Model entities and relationships explicitly. Each relationship carries source, time, confidence and status. New verified evidence updates the graph; conflicting evidence is shown rather than silently overwritten. AI can surface candidate connections or unusual changes, but the graph remains governed.

06

Example: bring it to life

Before visiting a farm, a field officer doesn't search five systems. They ask:

“What do we currently believe about this holding?”

The AI assembles:

14 parcels are associated with this holding.

3 active agreements apply.

Parcel 42 has conflicting boundary information.

Satellite imagery suggests a possible land-use change since the last inspection.

Natural England recorded an intervention nearby.

Two sources disagree about the hedge classification.

Confidence: moderate.

Crucially, AI doesn't pretend there is one truth when the evidence conflicts. It exposes what Defra believes, why it believes it, and how confident it is.

07

AI opportunity

AI can assist entity resolution, infer candidate relationships, detect conflicts, identify emerging signals and generate evidence-backed summaries. Every inferred relationship must retain provenance and confidence, with human/governed confirmation where it matters.

08

Potential value

  • A reusable intelligence layer across services
  • Faster detection of conflicting or changing evidence
  • Better operational and policy context
  • Foundation for Digital Twin and regulatory-network concepts

09

Reduce uncertainty without overcommitting

Smallest meaningful experiment

Create a graph for one narrow journey with 3–5 entity types and real source metadata. Compare task completion against the current search-and-join workflow.

10

Open questions and risks

What we still need to learn

  • Inference must never be mistaken for verified truth
  • Graph scope can grow uncontrollably
  • Needs clear stewardship, vocabulary and access control