Allocation-Level Contamination Measurement under Strategic Optimization in Retrieval-Augmented Generation

Sponsored retrieval turns a finite context window into a measurement problem: each query admits only a limited set of passages, and the allocation of that context changes the answer that is measured downstream. Once factuality or hallucination metrics influence allocation or pricing, they become strategic targets. A passage may survive lexical retrieval and still shift, contradict, or contaminate the final answer after the generator assembles context. We study sponsored retrieval as measurement under strategic optimization. The paper defines allocation-level contamination measurement for generator-mediated answers and instantiates it with an ensemble of natural-language-inference contradiction scores, unsupported-claim ratios, passage-set contradiction rates, citation checks, and faithfulness measures. This yields a report-independent scoring rule that can enter allocation decisions under exact allocation and a frozen estimator. A Vickrey-Clarke-Groves (VCG) mechanism provides one mechanism-design application of the measurement formulation. In a corpus-backed proof-of-concept evaluation spanning 2,637 passages from HotpotQA and TriviaQA, hybrid dense-sparse retrieval with cross-encoder reranking, and a DeepSeek-V4-Flash generator, the VCG instantiation reduces hallucination risk from 0.360 under plain VCG to 0.346 in the tested setting. The reduction is modest. Static controls remain strong baselines: heuristic caps and constrained optimization can reduce harmful sponsored exposure and, on some raw factuality and risk metrics, outperform the VCG instantiation. The empirical study provides proof-of-concept evidence for measurement-driven allocation under strategic content.