Useful AI knows what it doesn't know.
Every claim carries its state.
Noticed, not yet structured.
Structured claim with a named source.
Conflicting evidence. Both accounts kept visible.
Supported by evidence and validation.
Future state. Never mixed with current truth.
We do not use model confidence as proof.
Promote CLM-0311 “All supplier invoices are matched to POs” → Verified
model confidence: 0.94
Unchanged. Claim remains Unverified.
Where data can answer, data answers.
“Purchasing always gets the design pack from Asana.”
CLM-0142 · Purchasing.input ← Asana.design_pack
2 interviews · Asana export · 41 sampled tasks
Deterministic probe · 38 / 41 matched
Verified with known exception: urgent orders via email
Does project_id reliably connect Orders and Goods In?
Read-only by default
Investigation inspects systems. It does not change them.
Human approval
Promotions and actions that matter require a named person to approve.
Auditability
Every claim, change and decision keeps a path back to its source.
Current and future kept apart
Current State and Future State are separate models. Proposals never overwrite truth.
Discuss governance with our team.
We'd value your view on how your organisation needs AI evidence, approval and audit to work.