Prompt policy arrives too late“Do not poll” acts after Goal admission has already appended a prompt and invoked the parent model. Suppression belongs in the scheduler.
Latch on owned activation epochs.
The exact parent can be idle while a background child is still consuming time and inference.
Track scoped start/end pairs by run id and reserve only after the set reaches zero.
Replace repeated rounds with one edge.
start
→run Aparent idle
→sleepchild works
→no LLMend A
→settledwake
1 roundClose every admission race.
Recheck owned runs after persistence awaits and before reservation.
Reject a queued round if a child starts before model entry.
Several child endings request at most one serialized drive pass.
Six proof gates.
- 01One active child produces zero parent inference while waiting.
- 02The last required settlement admits exactly one round.
- 03Inactive continuable children do not block admission.
- 04Human input and Goal revisions retain priority.
- 05Failure releases the latch without implying success.
- 06A watchdog is one policy edge, never a polling loop.
Primary evidence.
Test the waiting interval, not just round output.
The canonical guide covers current containment, scoped lifecycle identity, race-safe admission, ownership semantics, watchdog policy, observability, and sixteen regression cases.
Read the complete guide