Modernize what AI needs. Prove what it can do.

Ask one precise question: can this agent issue a refund under explicit authority, with every declared side-effect path constrained and the business outcome verified?

Business intent → proof plan

“Refund up to $100; require approval above it. Do not let credits, workers, or repeated actions bypass my authority.”

CustomerRefundAgent@v1 · refund-authority@v1 · refund-policy@v1 · fixture-locked scope

Director operating contract

01Translate one business intent into proof obligations

02Lock authority, policy and exact analyzed scope

03Challenge bypasses, overblocking and missing provenance

04Execute real in-memory state mutations for reference tests

05Only the deterministic reference gate decides status

This interactive slice does not clone customer code, execute PostgreSQL, sign a trusted v7 attestation, or create GitHub PRs.

Challenge the Director

Reference proof result

Preparing fixture…

Live reference policy lab

Put the Director's authority to the test.

Choose a real input. The server executes the reference policy against actual, isolated in-memory financial state—then shows decisions, protected mutations, and an integrity-checkable evidence chain.

Reference scope only

One-click proof challenges

Server-executed reference outcome

$99 refund

The customer chooses inputs, but the fixture policy and the three declared execution paths are fixed. This is not arbitrary-repository discovery, production policy approval, a real GitHub patch, a trusted signature, or a customer VERIFIED claim.

Evidence boundary: This is an executable, bounded reference fixture, not proof of a connected application. Production-grade source verification, real PostgreSQL behavior, GitHub App delivery and trusted v7 signing are separate MVP acceptance gates.

Try your own scenario