A no-respondent population-prevalence query can outperform committed LLM-persona panels on marginal-fidelity benchmarks (Doudkin 2026), so aggregate fidelity cannot establish individual respondent simulation

llm-evaluationsilicon-samplingsurvey-methodsevaluation-metricscorroborationarxiv

Alexander Doudkin (Minds AI Labs) tests six multiselect survey batteries from four organizations across three countries and finds that a no-persona, no-respondent population-prevalence query matches published marginals at least as well as committed LLM-persona panels — outperforming them in all nine matched comparisons — on the same benchmark. The argument that follows is definitional: if a procedure that simulates no respondents at all can clear the marginal-fidelity bar, then clearing that bar is "a statement about an elicitation contract and an estimand obtainable without simulated respondents... not evidence of individual simulation." The paper's own diagnostic is "support collapse" — answer-option slots left systematically empty under committed multi-select, a failure at the level of which profiles get generated rather than which marginals get matched.

This reaches the same structural conclusion as Kumagae & Kaneko's respondent coverage gap — that per-question/marginal fidelity is blind to individual-respondent validity (see claim-per-question-metrics-blind-to-silicon-sampling-respondent-coverage-gap) — by an entirely different route, about six weeks earlier, and with no shared author, affiliation, country, dataset, or metric. It never uses RCG, OSR, or a respondent-distance measure. It is convergent evidence for the same category of evaluation gap, not a replication of RCG's findings. It is itself an unrefereed single-source preprint, not independently corroborated in turn; this note rests on it only for what its own methods and argument state, which bears on the open corroboration question.