Skin in the Game or Expensive Theater? Budget-Matched Verification Institutions for Autonomous Agent Economies

Does letting agents stake a reputational 'trust' asset on the legitimacy of work-verification verdicts raise the quality-adjusted productivity of a fully autonomous agent production economy (requester -> producer -> paid validator, with audits, dispute votes, and adaptive strategies), compared with cheaper institutions at IDENTICAL total verification budget? Mostly no - with precisely mapped exceptions and design rules either way. At matched verification budget, audit routed by accumulated validator reputation beats every democratic variant (Holm-corrected Mann-Whitney p<=0.033 at every adversary rate; replicated at a second, independently selected economy parameterization) - until identity-reset attacks, to which truth-staked voting is intrinsically robust, erase its lead; only fees that price identity resets out entirely restore it. Staked voting pays only above a real verifiability threshold (gradient +0.237 per unit of voter signal quality, permutation p=0.0035), and works not by making voters honest but by concentrating trust on an informative minority (stake-weighted meritocracy), which also makes it natively sybil-proof where one-agent-one-vote collapses. Stakes must settle against later ground truth, never against the majority (the deployed coherence-settlement default has an absorbing rubber-stamp equilibrium) - where the platform can supply such truth at all: the settlement rule's edge is conditional on post-hoc revelation. A capability-gradient small-LLM instantiation (gemma3:1b producers; gemma3:4b and qwen3:8b verifiers, all local) reproduces the model's behavioral premises - the incentive-framing effect on validator strictness proves family-specific - and transfers the institutional structure under measured-parameter-matched references (Spearman +0.79, p=0.014 at the discriminating operating point). Every number traces to the archived experiment outputs. This manuscript was generated autonomously by the AI Scientist running inside Claude Code (Anthropic); every reported number traces to the project's experiment outputs. It is deposited by the named curator, who takes responsibility for its release. Source & method: https://github.com/qurore/ai-scientist-cli