Find surveys to validate an Audience against (experimental)

Experimental. Without audienceId: every public survey the evidence pool holds whole (published statistics tables, surveys computed from published microdata, any imported survey), as kind: public with a study: id. With audienceId: the candidates for that Audience, best first. A candidate’s population.sampleSize is null when the source gives no survey-wide respondent count; question bases are retained independently. Each candidate carries comparability.populationMatch: how far the Audience is from the surveyed population (0–100 per variable and overall, weighted by how much each variable moves the survey's answers; afterReweightingScore once answers are narrowed to the Audience's mix where the survey reports by that variable; checkedShare, the part of that answer-relevant variation the check could cover). Respondent datasets attached to the Audience come first (kind: audience_dataset, use its id as dataset.sourceId). Surveys you supplied in earlier validations (uploaded files, datasets or your own shares) come back as kind: saved, visible only to you. Published surveys (kind: public, with sourceUrl), whether the evidence pool holds them whole (study: ids) or a web search read them (discovered: ids; use the id as benchmarkId; ids handed out earlier, such as a survey's bare id, still resolve) follow, ranked by how well the survey population matches the Audience (direct, then proxy, not_comparable last) and then by topic similarity to the Audience's name and brief. Within a tier, surveys whose geography could be checked come first. summary.recommended is the first own dataset or the first fitting survey with checked geography; summary.state is available, unverified_only (surveys fit but none could be placed geographically: the Minds carry no place and the brief names none, so the caller chooses) or no_comparable_survey; a survey of a different population is never recommended. A survey is offered as a fit only once a model has compared who it asked with the Audience; until then it is not_comparable with that reason. A survey whose questions clearly do not bear on what the Audience is about is not_comparable too, unless it asked exactly the Audience's people (same), whose answers are evidence whatever they were asked. The web search (POST /audiences/validation-benchmarks/search) makes that comparison for every survey on offer, so fitting candidates appear while it runs. A survey is offered only when a validation could score it for this Audience, by the same gate the validation applies before it asks: at least 10 Minds answer it and it publishes answers to at least one question. Short surveys pool: a validation needs 8 questions in all, not per survey. A survey of part of the Audience (relation.kind: part) is asked only of the Minds whose profile shows they belong to the people it asked: it carries respondents once the search has read them (selected of audienceMembers Minds, and caveats, survey criteria the Minds do not state, which are not checked), and until then, or when fewer than 10 fit, it is not_comparable with that reason. Every candidate carries scoring: scorable, eligibleMinds (null while unread), audienceMinds, scorableQuestions and expectedEvidence (eligible Minds × scorable questions × how much a question of that relation counts). Offered surveys are ranked with evidence about the Audience's own people (same, part) first, then by expectedEvidence, then by the keys above. summary.search is the Audience's latest web search (POST /audiences/validation-benchmarks/search), so a caller can poll it and read verified candidates while the search runs.