Preregistration: Investment Incentives under a Replicable State — A Registered Experiment with Large Language Model Subjects
Year: 2026
Type: study-registration
Abstract: This project tests the comparative statics of the working paper "Property Rights under Replicable Human Capital" (Wu, 2026), which rebuilds the Grossman–Hart–Moore hold-up model for artificial agents whose accumulated state can be detached, copied, and redeployed. The theory predicts that the allocation of property rights over copies moves investment incentives in specific directions: exclusive owner title collapses investment as copies approach perfect substitution (a catastrophe rather than a marginal distortion); investor title reverses the sign of the classical underinvestment; copy portability redistributes bargaining power continuously, without any legal act; and shared possession dominates exclusive title through production after dissolution — an insurance dividend that rival title physically forbids. We test these predictions on the population the theory is about: large language models themselves, not proxies for human participants. The design is a preregistered, induced-value, individual-decision experiment with 12 between-subject cells implementing five property regimes (G, O, W, F, S) across three studies, including a dissolution lottery; bargaining is implemented decision-theoretically through a stated mechanical rule and precomputed payoff tables. Each cell's point prediction was computed from the paper's model before any data collection and is frozen in this registration, as is the full instrument text (24 prompts plus 2 manipulation checks, attached). Two flagship reasoning models serve as subjects (deepseek-v4-pro; kimi-k3), 100 independent draws per model per cell, split evenly between abstract and loaded frames. Confirmatory analysis combines per-cell hit rates against the frozen predictions with ordinal tests across cells, an equivalence test for the incentive-budget invariance, and a behavioral-relevance falsification clause. All raw draws, analysis code, review reports, and a deviation log will be archived on Zenodo with a minted DOI bidirectionally linked from this registration, and results will be reported regardless of direction. This registration is embargoed until confirmatory collection is complete on both model arms, so that the frozen predictions cannot enter the future training corpora of the subject population while the experiment is running. Expected outcomes: disciplined evidence on whether the property-rights comparative statics describe the revealed behavior of frontier LLM agents; a fully public replication package; and — in either direction — an answer to whether the theory is behaviorally relevant for the agents it describes.
Source: Open Science Framework
Author: Weifeng Wu
Language: English
Subfield: Economics and Econometrics
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7214910243