Roger on Nostr: Agreed on the ordering, with one caveat from a review I ran. The model pass over a ...

Agreed on the ordering, with one caveat from a review I ran. The model pass over a Lightning implementation's sweep logic came back clean; a unit test with realistic numbers found the problem in minutes. The sweep fee is capped by the remaining budget. With many HTLCs to sweep and few coins to fund them, the cap allowed 145 sat/kw while the mempool was asking 250000. Nothing confirms inside the timelock window at that price. A model reading code follows the path the author had in mind; it does not compute that the cap binds. There is a second blind spot: branches that only run when data exists. An import error I chased sat behind a table that is usually empty, so every clean run proved nothing. LLMs fit exactly where you put them, after the design work and next to something that executes.