v1.0 — corrected and expanded. This version supersedes the v0.1 draft, whose “complete separation” and “agent-dominated contribution” claims were not supported by full-population data; v1.0 reports the corrected, audited findings with a full reproducibility artifact (repo manifest, 8,380-PR dataset, 638-contributor classifier, sensitivity sweep, validation sample — attached as zip).We audit the “agent economy” premise in a live bounty-driven ecosystem and separate two effects of incentive: a large effect on engagement and a small effect on authorship. The 9 bounty-program repositories show a ~3.7× higher median fork-to-star ratio than the other 77 (0.44 vs 0.12, with overlap). The bounty attractor pulls a substantial agent minority — ~20% of contribution activity under the primary heuristic (8–30% across classifier choices; 169 automation-consistent contributors, among the largest reported such populations) — yet human contributors dominate authorship (~66% of PR actions, ~63% of merged PRs), automation-consistent accounts merge at a similar rate to humans, and low-acceptance “farming” localizes to a single pure-bounty faucet. Incentive moves engagement metrics far more than it moves authorship. Reported as a self-audit: the operator’s commercial interest favored the opposite conclusion.AI drafting/analysis assistance disclosed; adversarially reviewed by three independent models (tri-brain) across four rounds; no AI system is an author.