Skip to content

RegBench — a benchmark for regulatory reasoning.

Compliance questions don’t have one string-matchable right answer. RegBench measures the thing that actually matters to a compliance team: whether a model finds the obligation that applies to them, reads it against their own policies, and shows its working.

What it measures

01

Retrieval under jurisdiction

Given a question from a firm in one jurisdiction, does the model reach for the instrument that actually binds them — or the better-known one from somewhere else?

02

Grounded interpretation

Read against a real policy set, does the answer follow the firm’s own controls and operating model, or drift into generic best practice?

03

Evidence that survives review

Every answer is graded on whether its citations hold up: right instrument, right provision, quoted as written, and traceable back to the source.

First results publishing soon

We’re running the first set now.

The method, the question set and the full results go up here when the first run closes. If you want them as they land — or want to see how Amoreg scores against your own regulators — get in touch.