How an estimate is made

The estimator

HiveSight does not simulate individual fake respondents and count their answers. Research consistently finds that persona roleplay compresses within-group variance, dramatizes hardship, and drifts with prompt wording — and our own benchmark measured it at 37 points of topline error. Instead, HiveSight treats the language model as a conditional response model over population cells:

1. Audience resolution. Your geography and filters select records from a calibrated synthetic population built from Census Bureau Current Population Survey microdata — 4.2 million person records with calibrated weights and state and congressional-district assignment.
2. Cell partition. The audience is partitioned into roughly 150 post-stratification cells by age band, earned-income band, sex, and — where a cell carries enough population weight — housing tenure, children at home, means-tested benefit receipt, and Social Security receipt.
3. Distribution elicitation. For each cell, the model is asked — as an expert estimator, not a roleplayer — for the percentage of that group choosing each response option. One call per cell, structured JSON out.
4. Weighted aggregation. Cell distributions combine with calibrated population weights — the same post-stratification arithmetic survey statisticians use. Subgroup estimates re-aggregate the same cells; nothing is re-asked, so breakdowns are always consistent with the topline.

What the verbatims are

Alongside the estimate, HiveSight writes a handful of illustrative verbatims for weighted-sampled population profiles. They are synthetic, clearly labeled, and carry no weight in the numbers. No human respondents are surveyed anywhere in this product.

Known limits

• Toplines on self-report scales carry a measured level bias (about 17 points MAE on our SHED suite versus 6 points on subgroups). Estimates are most reliable for comparisons — between segments, geographies, or message variants — where level bias cancels.
• A calibration layer that corrects levels against anchor surveys (GSS, SHED) is in development; it ships only when it passes held-out-question validation, and results will disclose when it is applied.
• The model is frozen in time: it cannot track opinion shifts after its training cutoff. Fast-moving topics degrade accuracy.
• Small or marginalized subgroups are where language models are least reliable. Cell estimates for narrow audiences deserve extra skepticism, and fielding real respondents remains the standard.

Intended use

HiveSight is pre-field research infrastructure: form priors, triage which audiences to actually poll, stress-test message variants across geographies you could never afford to field. Professional standards bodies (AAPOR, ESOMAR) endorse exactly this diagnostic use of simulated respondents and caution against substituting them for measurement of real populations. We agree, and the product says so wherever numbers appear. See benchmarks for current measured accuracy.

Reproducibility

Every run records its engine version, prompt version, dataset version, model, seed, cell count, and weighted audience — downloadable with the results. The benchmark harness and its raw call logs live in the repository, and the benchmark page renders the same artifact the engine is gated on.