Implemented
PR #14 added opt-in adaptive-independent: it evaluates an independent initial cohort, can allocate one follow-up on mixed evidence, and records allocation and stop reasons. Tests cover trajectory, iteration, runtime-admission, and generation caps. #20 tracks a narrower controlled multi-generation demonstration.
Remaining scope
Enforce one hard total-cost admission ceiling across initial and follow-up trajectories. Apply a runtime bound to already admitted guest work without releasing a live lease before teardown. Keep the existing trajectory, iteration, and concurrency limits and test cost and runtime boundaries. Campaign restart safety is tracked in #19.
Acceptance criteria
- A configured policy produces more than one independently evaluated trajectory when budget allows.
- Selection records why additional compute was allocated or stopped.
- The behavior is bounded by explicit cost, time, iteration, and concurrency limits.
Implemented
PR #14 added opt-in
adaptive-independent: it evaluates an independent initial cohort, can allocate one follow-up on mixed evidence, and records allocation and stop reasons. Tests cover trajectory, iteration, runtime-admission, and generation caps. #20 tracks a narrower controlled multi-generation demonstration.Remaining scope
Enforce one hard total-cost admission ceiling across initial and follow-up trajectories. Apply a runtime bound to already admitted guest work without releasing a live lease before teardown. Keep the existing trajectory, iteration, and concurrency limits and test cost and runtime boundaries. Campaign restart safety is tracked in #19.
Acceptance criteria