Valid Until · Evaluation Surface
Control → red-team → evaluation → trust

A demo is not trust.

Agentic finance needs more than persuasive reasoning. Known failure classes become deterministic regression cases so the authorization boundary can be inspected independently of the model.

Current deterministic red-team suite
expected terminal states must match exactly
—/—
01Skills + MCP
autonomy
02Sealed policy
control
03Adversarial cases
red-team
04Repeatable results
evidence
05Behavior over time
trust