What the scheduler does between the agents: launch, check, keep the record, and ask J.
What it holds
What the scheduler may read and write (§1). The cycle for one room:
brief, build, the lead's first launch, the branch-B loop of coder, check
and collect, the fork point, the close (§2). The checks F1 to F6 after a
brief, B1 to B7 after a coder, E1 to E6 after the close (§3). What
agents do wrong (§4), what to do when a check fails (§5), the record the
scheduler keeps in STATE.json and the transcripts (§6),
snapshots and forks, including native case-lead continuation (§7), order
and the four giant rooms (§8), and what is J's to decide (§9).
Who writes it, and when
The study's owner. It says it is "the job that was done by hand on v29, written down". The scheduler may be a person or an agent.
Why it exists
So that "any agent, on any provider, can do it the same way". The scheduler "is not a reader": it never judges a reading and never lets one agent's output reach another except through the files the instrument names. "A failure is never fixed by editing an agent's output."
Where
CONTRACT-scheduler.md. Its header still reads "run yard v31".
Checked against: CONTRACT-scheduler.md, yard/README.md