The Human Proxy Reliance Boundary: A Prospective Claim-Admissibility Protocol for Inferences from Simulated Human Behavior

Large language models are increasingly used as substitutes for human participants in interactive agent benchmarks, survey research, behavioral experiments, and product evaluation. Existing research shows that the validity of these substitutes varies by task, population, measure, and simulation configuration. The scientific question is therefore not whether an artificial proxy resembles a human in general, but whether evidence obtained with that proxy can support a particular claim about a specified human population under a declared tolerance and intended use. We propose the Human Proxy Claim-Admissibility Contract (HP-CAC), a prospective research protocol that separates proxy fidelity, claim validity, and decision admissibility. A contract specifies the target estimand, population, interaction regime, decision context, tolerable inferential error, protected subgroups, and validity horizon. A precommitted qualification procedure uses a limited human calibration sample, frozen proxy configurations, uncertainty-aware claim-specific tests, and a strictly independent human confirmation sample. Its outputs are ADMIT, HOLD, or REJECT within a bounded evidentiary scope; none represents institutional authorization to deploy an AI system. We specify testable hypotheses, estimands, an evidence firewall, transport restrictions, selective-risk and coverage metrics, strong composite comparators, and explicit failure conditions. The contribution sought is a reproducible way to test whether prospective claim qualification improves the safety–coverage–cost frontier beyond existing Sim2Real and statistical-validation methods. No implementation, experiment, measured performance improvement, general guarantee, or universal claim of novelty is reported.