In regulated finance, a mature AI system is judged as much by its restraint as by its capability. This console runs 20,000 synthetic bank decisions through a governance gate: stakes tiers, prohibitions, abstention when the data can't be trusted, and a breaker that suspends autonomy when results degrade. Change the controls and the whole stream re-runs in your browser.
Start from a strategy, then switch layers off to see what each one is worth.
Each column is a slice of consecutive decisions, split by verdict. Partway through, a drift episode hits card fraud and credit limits: the model stays confident while its accuracy collapses. Watch what the gate does with it.
Most AI assurance asks how often the model was right when it acted. This asks the opposite: when the system held back, was it right to? Each rule is costed: losses it averted against the review effort it spent.
| Rule | Cases held | Would have erred | Loss averted | Review spent | Net value |
|---|
of every decision where autonomy would have done damage, the system declined to act on.
harmful autonomous actions got through. This is the residual risk the policy owner signs for.
Every verdict is written to a hash-chained ledger: each entry's SHA-256 covers its record and the entry before it. Edit one past decision and every hash after it stops matching.
Every proposal is checked in this order. The first rule that applies decides, so the most restrictive verdict always wins.