Verifier-Bound Mechanism Discovery: Pre-Registered LLM Proposers for Mobile Crowdsensing Incentive Design
Description
Incentive mechanisms for mobile crowdsensing (MCS) are judged by two programmatic gold-standard properties: individual rationality (IR) and incentive compatibility (IC). LLM-based rule discovery has transformed combinatorial optimization and scientific discovery, yet a full-text arXiv sweep finds zero works applying LLM proposers to MCS incentive design, and the closest LLM-discovery systems accept proposals without programmatic property verification. We port a verifier-bound discovery engine (companion paper) to MCS mechanism design: an LLM proposes payment functions as sandboxed programs, an enumeration-based verifier checks IR, $ $-IC, boundedness, and budget feasibility on a type grid, and a 100-pair distribution-level protocol with pre-registered acceptance criteria compares LLM proposers against same-budget random controls and three human-designed baselines under a 20-evaluation budget. The verifier intercepted an infeasible parameter regime before any data was collected (an IR–budget incompatibility), demonstrating design-stage value. Results: under the frozen protocol the LLM produced only 3 verifier-passing mechanisms in 2,000 proposals (0.15\ rate $W = 1/100$ –- a pre-registered negative result. The gatekeeper analysis is the substantive contribution: all three human-designed reference mechanisms also violate at least one gold-standard property, and among LLM candidates that reached enumeration ($n=1,448$), IR–IC joint violations dominate (429, 29.6\ mode precisely. We argue the conjunctive verifier itself is a measure of domain difficulty: the same engine that yields a 55\ evaluation ledgers, and negative results are available from the authors upon reasonable request via email.