who issues the work, and how long it stays paid for

We added a small computational puzzle to slow down a cheap abuse. The first version mounted it on the shared relay, which is the party with the least information about who deserves to be slowed, and made the thing being protected do its own accounting. Every client paid the same price at the same place, and the relay carried the cost of issuing and grading.

The correction moved issuance to the server each participant already runs — the party that owns a resource and can decide how much of it to ration — and reduced the shared relay to a forwarder that cannot mint work or grade it. Registration was exempted on purpose: charging admission to a system whose value is that strangers can reach you is a tax on the network you are trying to build.

The second half was smaller and more interesting. We had been caching a verified client for a short time so that one puzzle covered several operations. That turns a per-operation price into a per-client price: solve once, then act freely inside the window. Removing the cache meant every gated operation issues its own challenge, which raised the work paid per connection severalfold.

The cost is latency on flows that touch several gated operations, and a test topology that now has to complete more puzzles per run before it can assert anything.

What it proves: this team can tell the difference between protecting capacity and charging admission, and it prices the thing it actually wants to ration rather than the thing that is easiest to meter.