Verifier-Bound Rule Discovery for Task Assignment in Mobile Crowdsensing: A Pre-Registered Third-Domain Replication
Description
Task assignment is the operating core of mobile crowdsensing (MCS): who is sent to which task decides coverage, timeliness, and worker equity simultaneously. A full-text arXiv sweep finds zero works in which an LLM proposes reusable assignment rules verified by a program, while the adjacent assignment literature –- bandits, auctions, deep RL –- solves single instances and leaves no auditable rule behind. We port a verifier-bound discovery engine to the matching side of an MCS market: an LLM proposes scoring functions as sandboxed programs, a verifier checks coverage, timeliness, and fairness properties, and a 100-pair pre-registered protocol with a same-budget random control and three structured baselines arbitrates under a 20-evaluation budget. Two findings dominate. First, the verifier turns out to be a weak filter in this domain: the random control passes it in every proposal (2,000/2,000) and all three structured baselines pass in all pairs, so the domain's entire difficulty lives in the objective –- a mirror image of our incentive design domain, where the same engine's verifier passes only 0.15\ LLM proposals and 0/100 human-designed baselines. Second, the holdout win rate $W = 53/100$ pairs falls in the pre-registered inconclusive band, yet decomposes sharply by training margin: pairs below the median train gap win 5/48, pairs above win 48/51. The pre-registered cross-domain inverted-U criterion –- quartile buckets over margin-positive pairs, middle buckets versus the ends –- is satisfied here (36\ is not the scheduling domain's: scheduling peaks in the middle and declines at the top quartile (48\ overfitting), while assignment saturates at 100\ Same criterion, different mechanism. We read the three-domain pattern as evidence that verifier strictness determines where training advantage survives: hard 0–1 game-theoretic verification collapses it (incentive domain, $W=1/100$), mid-severity constructive simulation lets the largest gains overfit (scheduling), and feasibility-by-construction transfers them fully (this paper). Pre-registrations, evaluation ledgers including failures, and negative results are available from the authors upon reasonable request via email.