Ariadne

A score is not a reason.

Auditable AML verdicts — the model stays out of the decision path. A BLOCK only happens when the model's risk band is corroborated by at least one explainable rule, every verdict is sealed in a tamper-evident chain, and written back to DataHub's ML lineage graph.

The corroboration gate

Move the risk band and the out-degree, or pick a preset. The verdict updates live — nothing here is precomputed.

The model's discretized band. A float never enters the decision.

Seal + verify

Sealing writes to the tamper-evident SQLite ledger AND a real INFERRED assertion to DataHub — this is a live write, not a preview.

Ledger chain

Each entry hashes its own payload plus the previous entry's hash. Altering any entry breaks every hash after it.

Day two: silent model redeploy

Reads the model's registered baseline for real from DataHub, then compares it against a deployed snapshot — catching a silent swap before it costs money.