Model governance

Bounded, versioned, reviewable.

No statistical model is in the decision path

The current engine is a deterministic rule set. There is no trained scorecard, no machine-learning classifier and no language model in the decision path. Nothing infers a borrower figure that was not supplied.

Versioning

Every run records the policy version and the engine version that produced it. Active: policy aeax-policy-2026.08.1, engine aeax-engine-1.2.0. Changing a threshold requires a new policy version; historical runs keep their original version.

Reproducibility

Runs store an input fingerprint computed from the normalized inputs and both versions. Re-running the same file under the same policy must produce the same fingerprint and the same outcome; a mismatch is a defect.

Data provenance

Each run is labelled by source type: live public data, borrower-entered application, or synthetic fixture. These categories are never blended, and synthetic fixtures carry no real identifiers.

Data completeness and the shadow rule

If any required credit-file input is absent, the run is labelled a shadow decision, the missing fields are listed explicitly, and the outcome is forced to human review. An incomplete file can never be auto-approved or auto-declined.

Human override boundary

Staff reviewers can move a file through the workflow and record notes. Those actions are written to an append-only event trail with the actor and timestamp. Reviewers do not silently rewrite a prior decision run.

Weight provenance

Sector and geography adjustments are configurable demo policy weights chosen for demonstration. They are not derived from any portfolio performance data, and we do not represent them as such.

Where AI may be added later

Document intake and field extraction are candidate areas for AI assistance, with extracted values shown for confirmation and marked as extracted in the audit record. Policy execution itself stays deterministic.

Fair-lending posture

The engine consumes cash-flow, trading history, credit-file quality, sector and country. It does not consume protected characteristics. As a prototype it has not undergone disparate-impact testing; that testing is required before any production use.