How We Adapted Simula for Compliance Scenarios

We built a synthetic compliance scenario generator for our stablecoin platform. The rule packs were not the main problem. The system needed reproducible and auditable test data. The data also needed enough variation to stress each rule pack.

We adapted an existing method. Davidson, Seguin, Bacis, Ilharco, and Harkous describe it in Reasoning-Driven Synthetic Data Generation and Evaluation. Transactions on Machine Learning Research published the paper in March 2026. The paper introduces Simula. Simula maps a target domain into explicit taxonomies. A generator and critic then use those taxonomies to produce varied, reproducible examples.

We mapped financial-crime typologies to Simula’s taxonomy nodes. Examples include layered remittance, proximity to sanctioned counterparties, and reserve-drift patterns. The generator creates synthetic counterparty graphs and transaction sequences. The critic rejects a scenario when it does not contain the required typology.

What we added

Simula provides taxonomy expansion, generator refinement, and critic-based quality checks. We added three controls for regulatory review:

1. Provenance fingerprint. Each scenario contains a hash of its manifest, rule-pack version, and assumptions. A reviewer can use the hash to compare runs.
2. Append-only audit ledger. Each dossier export adds a row to an organization-scoped Supabase table. The database blocks `UPDATE` and `DELETE` operations. Reviewers can inspect the complete run history.
3. Synthetic-only enforcement. The schema rejects records that combine synthetic scenarios with live customer data. This control keeps the test environment separate from production data.

A limit of the method

A clear taxonomy cannot correct weak model reasoning. Davidson and the other authors tested Simula across multiple datasets. A regulated deployment needs equivalent tests for its own domain. We therefore run engine-disposition scoring in a separate evaluation harness. The first live corridor will compare synthetic scenarios with real pilot data. Synthetic data will not replace the pilot data.