Running is not the same as working
We track what we ship. We rarely audit what we assumed was working and never checked again.
There is a specific discovery you make when you dig into something old: it was never working. Not recently broken. Never worked. Running, scheduled, accounted for in your head as a thing that exists - and zero output since the day you built it.
I found this yesterday. A set of automated notifications on one of the businesses I run, scheduled to fire since late March. Quietly skipped 519 times in a row. Not one message sent. The scheduler was running. The endpoints existed. On my internal map, that area was covered.
It wasn't covered. It had the appearance of coverage.
The same afternoon, in another part of the operation, I found an AI assistant answering customer questions from a knowledge base that hadn't been touched in months. Retired course names. Stale dates. Changed prices. The responder was confident. The knowledge was wrong. Nothing in the system flagged it, because the system was running fine.
The category is the trap
As a solo operator, you build a mental map of what's handled. This is necessary. You cannot hold everything in working memory, so you compress. "Notifications: done." "Responder: live." The compression is right - you would drown otherwise.
But compression loses texture. "Done" starts to mean "exists" rather than "works." And because it exists, you stop asking whether it works. You stop asking because the asking is itself work, and you are already at capacity, and everything looks fine from the outside.
The cost is small per item and invisible. It compounds in corners. You end up with infrastructure that has never fired, knowledge bases quietly feeding wrong answers to real customers, staging environments for finished projects still open while the internet crawls them. None of it is a crisis. All of it is drag.
"Real artists ship" - and that's true, and it's half the story. The other half is that real operators retire. They close the door on things that have no job left. They trim the knowledge base. They retire the endpoint. They set the staging site to closed and move on.
I shipped 519 scheduler runs on those endpoints and got nothing out. Retiring them took five minutes. That ratio is the thing worth sitting with today.