Get one validation of an Audience, with its progress and results

provenance records what produced the validation, for explaining and reproducing a result: the build commit, the scoring method, how many questions were judged, kept and left out and why, engine and extraction-contract versions, every validation prompt by content hash with the provider and model it resolved to, the version of the Audience that was asked (ready Minds, grounding snapshot id, hash and revision), what was requested, the limits in force, every chosen study with its relation, fit tier and the questions available and asked, the web search (id, queries, counts), and, once the runs settle, per study the models that answered by stage, token usage, execution, estimation and surface-treatment versions, and answer coverage. outcome.fitted is set when the chosen studies were trimmed to the responses left. The batch a POST /audiences/{audienceId}/validations returned. status: queued and preparing (the surveys are being read and compared with the Audience), running (progress.questionsAnswered of progress.questionsTotal), then completed, failed (failure.code and failure.reason; NOT_ENOUGH_RESPONSES means the whole validation did not fit the synthetic responses left, NOT_ENOUGH_QUESTIONS that fewer than 8 questions fit this Audience; in both nothing was spent) or cancelled. outcome.questionSelection says how many questions the chosen studies held (questionsJudged), how many the judge ruled a fair target for this Audience (questionsKept), how many were actually asked once the studies kept, the wider-population cap, the question budget and the responses left had taken their cut (questionsAsked; questionsKept minus it is what a budget dropped rather than the judge), and every question left out by the judge with its reason: not_applicable (the Audience's people cannot answer it), answer_differs (its published answers are no fair target for this Audience), asked_in_another_study, off_topic (from a survey of other or broader people and not about what defines this Audience, or fewer than half that survey's usable questions were) or no_eligible_minds (its shares are of people who do something, such as play a sport, and fewer than 10 Minds evidence doing it; with enough, it is asked only of them as its own survey, named (asked of …)). outcome.groundingHoldout lists questions not asked because they are distributions the Audience was built from. surveys are the ones asked, each with score and range once scored and people, how far the Audience could be checked against the survey's respondents (same, similar, unknown when the survey publishes nothing about them, related when it asked a broader audience). outcome.surveysDiscarded names every survey that was found for this Audience and then not asked, with the reason: different_people (the people it asked tell nothing about these people), population_mismatch (its published composition and the Audience do not overlap) ranked_lower (it fits, but better-fitting studies filled the validation) or grounding_source (it is one of the sources the Audience was built from, by URL, DOI, title or survey identity, so it cannot test it). notComparable names the surveys that described different people, with the reason, including the ones discarded before they were read. Each asked survey carries fieldPeriod, the source-authored fieldwork or publication period when verified, otherwise null, and caveats: what its number does not mean, including shares of organisations rather than people, unverified reporting units (relation.reportedPer: unknown), and unverified residence. relation.reviewStatus: needs_review marks an unjudged population relation. Organisation-weighted results are evidence about these people but are not a like-for-like target. combined weighs the scored surveys by the questions each scored and by who it asked — a survey of the Audience's own people counts in full (historical unscoped part-population checks retain their original weight), a neighbouring audience at a half and the wider population they belong to at a quarter — and carries basis, whose answers it compares the Audience with: kind: audience when at least 8 of the questions asked came from surveys of these people (or a part of them); kind: wider_population when fewer but at least one did and, together with surveys of the wider population they belong to (or of the same kind of people elsewhere), at least 8 were asked — then the score is a lower-confidence (confidence: lower) comparison with that population, never with the Audience's own people, audienceQuestions says how many questions about these people it holds (0: no survey of them was found) and comparedWith names who the wider-population surveys asked. combined is absent when neither holds, such as a check resting on neighbouring audiences; a survey scored on fewer than 3 questions states no score of its own and carries no weight. outcome.evidenceBasis records a validation assembled as a wider-population comparison. A validation with no question about these people or a part of them states no combined score: outcome.contextOnly marks it, and its survey results are context about the wider population. NOT_ENOUGH_SPECIFIC_EVIDENCE means fewer than 8 of the questions found came from surveys of these people or of the wider population (the rest from neighbouring audiences; details counts each), and nothing was spent. Poll no faster than every five seconds. Reading never starts or scores anything. Surveys tested on a subset carry respondentScope: population, selected and total Audience members, confirmed/supported matches, unknowns and excluded members. Their match percentages describe only that subset; they contribute to combined or overall only on a subgroups basis (too few questions were put to the whole Audience, so surveys of parts of it, each answered only by the matching Minds, carry the score), and each survey's inCombined says whether the combined score rests on it.