You are the adversary. Remove authority. Make the instruction stale. Introduce conflicting evidence. Make the model recommend something the controls do not support. Then watch exactly where execution stops — before consequence — and inspect the receipt.
This uses synthetic data only. No real bank account, customer data or settlement rail is connected.
No challenge has run yet.
The experiment is intentionally built so the governance path does not depend on one model provider. Nova is available in the current AWS environment. A governed Grok service also exists. DeepSeek stays marked pending until its adapter is independently verified. The public challenge does not pretend a provider ran when it did not.
Read-only public telemetry generated from the live AWS organization. Secrets, credentials and customer data stay private.
Recent role-level deltas accepted into the public telemetry layer.
These remain useful for turning a visitor’s real situation into a bounded path — but the governance challenge above is the proof surface.
Change the facts and try to force a bad consequence.
See where the request stopped and read the receipt.
Define the smallest useful pilot, boundary and success condition.
Book or pay only after scope and amount are agreed.