Public AI benchmark scores are contextual signals, not deployment evidence. A simulation finds that changing benchmark holdouts can alter the relationship between published scores and modeled user satisfaction, but it does not show how that relationship behaves in enterprise deployments.
In the modeled ecosystem, moving from public-only to private-only holdouts narrowed the score–satisfaction gap on most benchmarks and widened it on some. The widening depended on whether holdout weighting shifted scoring credit toward capability dimensions where provider R&D was concentrated, and on the composition of the benchmark suite.
Those results come from a stylized hybrid agent-based simulation of providers, evaluators, consumers, regulators, funders and media. The authors describe its outputs as qualitative dynamics, directional patterns, effect ordering and mechanism conditions rather than point predictions about real markets.
The study does not establish whether enterprise proofs of concept dampen or amplify the modeled relationship. It does not estimate the correlation between buyer-side, deployment-specific benchmarks and public benchmarks, and it does not model internal use-case-specific evaluation suites trusted above public leaderboards. It also provides no evidence of binding legal duties, deadlines, enforcement authority or penalties.