The Axiomyx Platform builds the evidence ledger before AI writes a word — canonical payload, β/Df/Dh signature, Zone v2 placement, graph path, and reproducibility hash. Generated narrative is shown separately from those deterministic outputs and remains subject to human compliance review. This demonstration does not establish production detection performance.
Many AML AI workflows depend on probabilistic scoring, embeddings, or free-form narrative generation. That creates governance pressure: investigators need to know why a transaction was flagged, what evidence was linked, and whether the result can be reproduced.
Free-form narrative can introduce relationships that are not present in source evidence. Reviewers need a clear boundary between generated prose and recorded evidence.
Model wording and interpretation can vary. Reproducible evidence fields make it possible to compare the generated narrative with the same recorded inputs.
An alert without a source and transformation path is harder to examine. Graph Provenance demonstrates an inspectable route for this synthetic sample.
Click a synthetic transaction and follow it from raw record to deterministic evidence outputs and a separately labelled AI-generated investigation narrative. The result remains a research demonstration, not a regulatory decision.
Dataset: IBM public synthetic AML transaction data (IBM/AML-Data). CSV financial transactions with laundering tags used for benchmark evaluation. Repository licence: Apache-2.0. Data licence: CDLA-Sharing-1.0. Independent Axiomyx benchmark — no IBM endorsement implied.
Displayed transactions pass through the five-component pipeline. Canonical fields, geometric placement, relationship outputs and provenance are deterministic demo evidence; the AI investigation narrative is generated prose that requires review.
🔴 Suspicious cluster 🟡 Under review 🔵 Normal pattern
Each displayed transaction and deterministic component output is timestamped beside the generated narrative. The evidence fields can be replayed; the model prose is not represented as deterministic truth or a regulatory filing.
| Timestamp | Transaction ID | Amount | β / Df / Dh | Zone | Entity Match | Risk | SAR Narrative |
|---|
This working reference shows five deterministic components processing synthetic transactions into inspectable evidence before AI writes a word. It does not prove live detection performance, eliminate governance work or replace a qualified compliance decision.