Learned orchestration: the policy and the critical path

Budget-run lab (Module 9, L9.6, advanced): the spawn count is a decision, not a constant, and the latency it buys is the longest path through the task, not the number of steps anyone ran. Seven spawn policies are priced over the same 12 tasks of difficulty 1-5 under one $1.00 budget at $0.02 per worker per tick. No fixed fan-out both fits the budget and clears every task: fixed-8 succeeds at 12 of 12 for $1.92 (OVER, $0.160 per success), fixed-5 is the cheapest fixed plan that clears all 12 at $1.20, fixed-4 is the largest that fits at $0.96 and clears only 7 of 12, and fixed-2 fits at $0.48 and clears 3 of 12. One statement beats them all -- spawn the estimated difficulty: $0.94 at 12 of 12, $0.078 per success, 2.05x less per success than fixed-8 -- and guessing 4 on every task costs $0.96, $0.02 more, at 7 of 12, so a bad estimate is not even the thrifty option. Part 2 is the metric: the task is a four-step DAG (research 3, write 6, test-gen 2, review 2 ticks) whose baseline latency is 11 ticks via research -> write -> review. 2x workers on test-gen buys 0 ticks for 1 more worker ($1.10 against $0.88), 2x on write buys 3 ticks (8 ticks), and at 6x on write the path flips to research -> test-gen -> review at 7 ticks, where the next dollar belongs on test-gen. Every allocation runs the same 4 steps and the same 13 ticks of total work, so neither count nor work is the clock. ADVANCED: the training loop itself is a later module's subject; this lesson teaches the framing, the metric and a heuristic policy. In-memory only; scripted tasks and a scripted DAG, illustrative dollars, no threads or provider call.

This Harness Engineering lesson is part of a paid plan. Sign in with a plan that includes this course to read it.