Settlement is not instant, but the request is waiting
A demo settles in a friendly network in a second. Under load, confirmation is a distribution with a long tail, and your HTTP request is holding a connection open the whole time. You either block — and inherit the tail as your p99 — or you return early and reconcile later, which means the “retry with proof” step is now asynchronous and needs its own state.
Refunds run the handshake backwards
The moment you charge, someone pays for something that then fails to provision. Now you owe a refund, which is a settlement in the other direction, against a record of what was bought. This is why the provisioned-record is not optional: without it, a refund has nothing to reverse and becomes a support ticket.
Disputes have no card network behind them
Card payments come with a chargeback process you did not have to build. Machine settlement typically does not. If a client claims it paid and got nothing, the arbiter is your logs and the chain, and the policy is one you wrote. That policy is a product decision disguised as an edge case.
You now depend on a facilitator
The facilitator is the piece that keeps your Worker off the chain, and it is also a new single point of failure and a new trust assumption. Its downtime is your downtime; its fee is your margin; its settlement claim is one you are taking on faith unless you verify independently. Choosing it is the most consequential decision in the stack.
Pricing drift
A price you hard-code is wrong the moment the asset moves or your cost changes. Production x402 pushes the amount into the challenge so it can float, which means the amount is now a live configuration value, not a constant — and a mispriced challenge is a request that either never pays or overpays.
Why this deployment sidesteps all of it
Every row above is a cost of charging. A deployment with no marginal cost to recover pays those costs for nothing. Here the equivalent concerns — abuse of an open endpoint, pages piling up forever — are handled by a rate limit and a 24-hour default expiry, both free and neither on call at 3am.
curl -X POST https://read.botook.ai/v1/publish -F file=@index.htmlIs asynchronous settlement really necessary?
Only under load. If your volume is low and your facilitator is fast, blocking is fine. The moment the tail of confirmation latency shows up in your p99, you will want the retry to be asynchronous, and that is a larger change than it looks.
How do refunds interact with rate limiting?
They are separate ledgers. A refund reverses a settlement; a rate limit throttles requests. Conflating them — refunding by resetting a counter — loses money the first time someone notices.