The spatial model of vote choice is the standard account of how Americans vote, and no dataset can fit it properly to an American election. The instruments that ask a voter to place himself and the candidates on the same policy scales are national and cover the presidency alone. The instrument with samples large enough for a single race carries one left-right number and places

one of the two people on the ballot. None of them asks a voter how much any issue weighs in his own vote, so the weights have to be read off the vote the model exists to explain. We build the missing data for one congressional race. A panel of 4,865 language-model respondents, each seeded on a real survey respondent’s recorded answers joined to a voter-file record and drawn so that the panel matches the district’s age, sex and county margins, answers a frozen interview of eleven question blocks. It asks where he stands on three scales and 32 issues, how much each of them matters to him, where he places both candidates on all of them, how he rates each man on competence and honesty with policy set aside, and, in a separate conversation, whether he will vote and for whom. That comes to 144 numeric answers before the vote is reached, which is why no human panel has been asked for it. One pass over the whole panel cost 142 dollars and returned 4,672 valid answers.

We report the costs, which rules answers fail and how often, and what the answers look like. The candidate placements and quality ratings from this pass carry a disclosed assembly defect and are not used. We release the interview, the source of every item, the component loadings and the record schema. The panel measures a simulated population and does not substitute for asking people.