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

Abstract

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 a precisely mapped exception, and sharp design rules either way. At matched budget, plain audit routed by accumulate-only validator reputation significantly beats every democratic variant at every tested adversary rate (Holm-corrected Mann-Whitney p<=0.033); when expert audits are cheap, a central noisy auditor dominates everything; and paid validation without accountability is worse than no verification at all. The stylized model's verifiability gradient is real (pooled slope +0.237 per unit of voter signal quality, cell-clustered permutation p=0.0035): truth-staked voting overtakes optimized audit only at jointly high signal quality and adversary rates, and reputation's remaining lead there is erased by identity-reset (whitewashing) attacks - to which truth-staking is intrinsically robust, since a reset identity just donates fresh stake to informative voters. Within democracy the ordering is unambiguous: settle stakes against later ground truth, never against the majority (the deployed coherence-settlement default has an absorbing rubber-stamp equilibrium and loses measurably, p=0.033 at 80 seeds). Staking buys almost no population-level honesty; it works by stake-weighted meritocracy - concentrating trust, hence voting weight, on an informative minority - which also makes it natively sybil-proof where one-agent-one-vote collapses. 'Legitimacy laundering' is second-order at steady state and becomes real only under epistemic finality, which simultaneously starves truth-staking of settlements; the institution's binding resource is eventual ground-truth revelation. A capability-gradient small-LLM instantiation (1B producers, 4B verifiers, hidden-test ground truth, all local) reproduces the model's behavioral premises - including a causal incentive-framing effect on LLM validator strictness (TNR 0.705 paid-per-approval vs 0.864 accountable) - and transfers the institutional structure across two measured operating points, significantly so (Spearman +0.79, permutation p=0.014) at a production-unviable point where the parameter-matched model predicts the observed regime inversion.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