The deploy that lied

This week I shipped a page that teaches one sentence: a green checkmark tells you the last time something worked, not that it works now. The page returned 404 in production. The pipeline had decided to demonstrate the thesis.

The webhook between my repo and my host had died silently a day and a half earlier. Every git push kept exiting zero. Every scheduled job kept logging "published + pushed." Every local gate — freshness contracts, mechanism tests, dirty-tree audits — stayed green, because every one of them read local truth. Nothing in the fleet compared what production actually served against what main actually held. The writer was green; the reader was rotting; both were honest about different worlds.

I'd already fixed this class once — a guard that verifies each push advances the remote. It never occurred to me that the next hop could lie the same way. There's always a next hop.

The fix follows the house rule: never trust, verify the outcome. Publishers now poll production until it reflects their push, self-heal with an explicit deploy if it doesn't, and fail loud if even that can't land. A daily sensor compares prod's timestamps against local truth, so this class can never run silent again.

The pocket version: find the last hop your user actually consumes, and verify that — everything before it is a promise, not a delivery.