the coordination layer we had to build to run the work

The research programme runs many lines at once: several independent question tracks, each with its own tasks and its own worker, all feeding one set of records. The first attempt was optimistic — parallel workers in a shared tree, with conventions held in people's heads.

The failure arrived within a day, and it was not technical in origin. A worker was bootstrapped into the wrong session, so its commands addressed a default context nobody was watching, and its writes landed outside the scope the coordinator believed it had granted. Nothing was corrupted; a snapshot comparison proved that, which is the only reason this is a paragraph rather than an incident review.

What we built afterwards is now a set of artifacts. A mandate stating what the coordinator owns. A numbered decision register, so an argument already settled is not re-run by the next worker. A roster naming each line. A guard on the push path that stops a worker from writing to the main line. A separate checkout per line. And a rule that every command is addressed to the session it belongs to, never to whatever happens to be default.

The cost is that the coordination material is larger than the code of any single task, and it has to be kept true or it becomes the stale map problem in a new costume.

What it proves: this team treats the process that produces work as an artifact with guards and verification, not as a habit that happens to be working.