Pre-sign policy gate for mppx agents (tested charge adapter)

[mercadosclaro-eng]

We built a non-custodial authorization layer for agents that use paid tools. Before the wallet creates a payment credential, it evaluates the selected payment intent and returns one of three outcomes: ALLOW, BLOCK, or REQUIRE_APPROVAL.

For mppx, the reference adapter uses the current prepareRequest() boundary and calls PreparedRequest.payment.pay() only after a matching, unexpired ALLOW. The MPP account and private key stay inside mppx; the policy service never receives them.

What we verified:

• the official mppx@0.11.0 runtime creates no credential-bearing retry before ALLOW;
• BLOCK and REQUIRE_APPROVAL create no credential and no paid retry;
• timeout, network error, malformed response, receipt mismatch or expiry, incomplete challenge, and a challenge changed during authorization all fail closed.

Current scope is intentionally small: one-time charge only. Sessions and subscriptions are rejected, and no custody, signing, settlement, fraud guarantee, or official MPP endorsement is claimed.

• Adapter, tests and usage: https://github.com/mercadosclaro-eng/ai-payment-control-plane/tree/main/integrations/mppx
• Public integration guide: https://mercadosclaro-eng.github.io/ai-payment-control-plane/integration.html
• Independent evaluation kit: https://github.com/mercadosclaro-eng/ai-payment-control-plane/tree/main/integrations/evaluation

We are looking for one MPP team with a real agent-payment path to validate the integration boundary. The controlled pilot form requests only non-confidential workflow details: https://github.com/mercadosclaro-eng/ai-payment-control-plane/issues/new?template=pilot-request.yml

Technical feedback on the request binding and fail-closed behavior is welcome.