Evidence Packs
Audit-ready AI outputs. Every answer includes citations, provenance graphs, and decision traces – enabling decision-grade confidence across all our AI solutions.
“Truth will come to light; murder cannot be hid long.
You'll recognise this as a needed capability if…
OIA requests take days to assemble because there's no record of what the AI accessed or why
Audit teams can't verify how AI-assisted decisions were reached – there's no decision trace
Staff can't show a Minister or oversight body exactly which documents supported a conclusion
You've paused AI adoption because you can't prove outputs are defensible under scrutiny
What's in an Evidence Pack?
An Evidence Pack is the bundle returned with every AI output. It shows exactly why the result is correct and is designed for auditability.
Your Question
Natural language query
Knowledge Base
Documents & data
AI Processing
Retrieval & reasoning
Guardrails
Policy validation
Evidence Pack
Audit-ready output
Your Question
Natural language query
Knowledge Base
Documents & data
AI Processing
Retrieval & reasoning
Guardrails
Policy validation
Evidence Pack
Audit-ready output
Every Evidence Pack Contains:
Example Evidence Pack
{
"job_id": "ep-2024-08-15-0847",
"query": "What are the procurement thresholds for government contracts?",
"timestamp": "2024-08-15T14:32:07.421Z",
"answer": {
"summary": "Open tender is required for contracts exceeding $100,000...",
"policy_title": "Government Procurement Rules",
"effective_date": "2024-07-01",
"administering_agency": "MBIE"
},
"citations": [
{
"doc_id": "DOC-2024-0847",
"pages": "12-14",
"char_range": [3420, 3892],
"quote": "Contracts exceeding $100,000 require open tender...",
"confidence": 0.94
}
],
"provenance": {
"chunks_retrieved": 12,
"chunks_used": 3,
"graph_hops": 2,
"combined_confidence": 0.91
},
"guardrails": {
"policy_version": "gov-standard-v2.3",
"pii_check": "PASS",
"legal_advice_check": "PASS",
"speculation_check": "PASS",
"redactions": []
},
"execution": {
"tools": ["vector_search", "graph_query", "llm_summarise"],
"latency_ms": 1247,
"tokens_used": 3420,
"model": "gpt-4-turbo"
}
}
Five components of proof
Every Evidence Pack includes these five components, providing a complete audit trail from question to answer.
Answer Summary
Structured fields aligned to your use-case ontology. Clear, consistent outputs that match your business requirements.
policy_title: "Government Procurement Rules", effective_date: "2024-07-01", administering_agency: "MBIE", status: "Active"Citations
Document IDs, page and offset ranges, and verbatim quotes where required. Every claim traced back to its source.
DOC-2024-0847, pages 12-14, chars 3420-3892: "Contracts exceeding $100,000 require open tender unless..."Provenance Graph
Visual representation showing which chunks and records supported which claims, with confidence signals for each connection.
Query → chunk_847 (0.94) → claim_1 | chunk_312 (0.87) → claim_2 | Both claims → final_answer (combined: 0.91)Policy/Guardrail Results
Complete record of validators passed or failed, any redactions or refusals, and the policy version applied.
pii_check: PASS | legal_advice: PASS | speculation: PASS | policy_version: "gov-standard-v2.3"Execution Trace
Full log of tools called, timing information, cost telemetry, and job IDs for debugging and optimisation.
job_id: "ep-2024-08-15-0847" | tools: [vector_search, graph_query, llm_summarise] | latency: 1.2s | tokens: 3,420
Why Evidence Packs matter
In regulated environments, “the AI said so” isn't good enough. Evidence Packs give you the proof you need.
OIA Requests
How do you prove what information the AI accessed?
Every claim can be traced back to source documents with page numbers and exact quotes.
Ministerial Questions
How do you show the sources behind a conclusion?
No more "the AI said so". Show exactly which documents supported the conclusion.
Audit Review
How do you demonstrate decision trail compliance?
Every output is versioned, traceable, and auditable with complete provenance.
The non-negotiable
Evidence Packs are our non-negotiable. Any AI output that leaves the sandbox must have one, or it returns a refusal with a reason and policy version.
No exceptions. This is how we ensure every AI output is defensible, traceable, and audit-ready.
Evidence Packs in production
Every AI agent we deploy – across multiple government agencies – produces Evidence Packs as standard. This isn't a feature we're building. It's been in production since day one.
Every AI output in production includes a full Evidence Pack – citations, provenance, and guardrail results.
All running with Evidence Packs attached to every response. No exceptions, no opt-outs.
Zero AI outputs have been released without source attribution. The system refuses rather than guess.
