AI CHANGEEvidence Engine
Method & trust boundaries

Evidence first.
Human decision.

The AI Change Evidence Engine helps a reviewer understand one completed material change without reconstructing its evidence across disconnected systems.

It does not approve the change, replace source systems, prove control effectiveness, or determine APRA compliance.
Deterministic pipeline

From source event to reviewer trace

  1. 01

    Observe

    Receive a narrowly scoped event from an approved read-only source and retain only the metadata needed for review.

  2. 02

    Normalise

    Map source-specific fields into one evidence model without replacing the upstream system of record.

  3. 03

    Correlate

    Connect records using explicit change, commit, workflow, artefact and deployment identifiers.

  4. 04

    Evaluate

    Apply a versioned, bank-defined profile. Report matched, incomplete or attention—not compliance.

  5. 05

    Present

    Give the named reviewer provenance, limitations and a digest-bearing export for a human decision.

What the reviewer can trust

Visible provenance and explicit rules

  • Every matched record names its source and source reference.
  • Direct, derived and asserted lineage are distinguished.
  • Requirements and accepted statuses are versioned data.
  • Missing evidence stays unresolved and visible.
  • The representative export has a reproducible content digest.
Where the prototype stops

Authority remains outside the product

  • Source systems remain authoritative.
  • Delegated authority must be confirmed by the bank.
  • A successful tool result is not automatically an effective control.
  • A digest detects modification; it does not authenticate the source.
  • The final decision belongs to a named human reviewer.
Capability map

What exists now—and what does not

Reviewer evidence passportDemonstrated

Representative case and deterministic profile evaluation

GitHub-to-passport loopImplemented

Verified ingress, multi-fact normalisation, exact correlation and live read model

D1 persistence modelImplemented

Release cases, evidence keys, deliveries and evaluation snapshots

JSON integrity exportDemonstrated

Deterministic SHA-256 digest; deliberately unsigned

ServiceNow, scanner and monitoring connectorsRoadmap

Representative records only; no live customer connection

Production signing and transparency logRoadmap

Not represented as a current capability

Working source-linked demonstration — not a customer control decision

The live GitHub path is implemented. Customer use would still require approved connections, customer-defined policies, identity and access controls, operational monitoring and production key management.

Learn how this was engineered →