How it works
Anatomy of a decision run.
- 01
Ingest
A loan file arrives from a borrower form, an operator entry, an API, or a public listing. It is normalized into one schema and tagged with its source type and source record ID.
- 02
Apply policy
The active policy (aeax-policy-2026.08.1) runs on the engine (aeax-engine-1.2.0). Thresholds for coverage ratio, margin, debt-to-revenue, trading history and credit file quality are configuration, not code buried in a model.
- 03
Explain
Each threshold crossed or missed emits a reason code with a stable identifier, a human-readable label, a detail sentence and a signed points contribution.
- 04
Route
Complete files inside policy are pre-qualified. Files that miss a required input, or sit outside policy bounds, route to human review — the agent cannot widen its own authority.
- 05
Audit
The run is written append-only with the normalized inputs, the missing fields, both versions, an input fingerprint and a timestamp. Reruns link to their parent run.
What a complete credit file requires
Any run missing one of these is labelled a shadow decision and routed to review.
- Requested amount
- Monthly revenue
- Monthly operating expenses
- Existing debt service
- Months trading
- Credit file quality
- Sector
- Country
Determinism check
Every run produces an input fingerprint derived from the normalized inputs plus the policy and engine versions. Running the same file twice under the same policy must yield the same fingerprint and the same outcome. The live demo includes a “run again” comparison that shows this explicitly.
Run a live decision