Prefield runs discovery interviews, hypothesis checks and concept tests on simulated consumer panels — seeded from government microdata, calibrated to each market's benchmark surveys, and graded in public when the real numbers land. Minutes and dollars, not weeks and five figures.
Pay per studyNo subscriptionOne pack per market
— If your bank paid your wages two days early for $5 a month, would you switch?
“Two days early doesn't fix the math, it moves it. I'd take it free, sure. Paying five bucks to get my own money sooner — that feels like a fee for being broke.”
The category's standard proof is retrodiction — regenerating surveys whose answers were already public, then self-scoring the match. Impressive percentages, incommensurable methods, and a test the model may have already seen. Prefield is built in the opposite direction: the panel's forecast is sealed before the benchmark survey publishes, and the release itself does the grading.
Every market's panel starts as official microdata — census households, labor-force spines, income and values surveys — raked to published margins and joined into whole persons: age, region, household, income, work, party, priors. Source licenses are documented per pack.
The panel answers the same items as the market's anchor survey. Per-domain corrections are fitted, validated on held-out items, and refit on the anchor's own cadence. What doesn't fit is documented — not hidden.
Studies run in the language and register the market actually speaks — formal Filipino in Manila, English in Ohio, usted-register Spanish in Monterrey next. Long-form transcripts, quantified syntheses and verdicts come back in minutes, segmented by the demographics the seed carries.
Ahead of each anchor release, the panel's forecast is frozen: SHA-256 published, distributions locked. When the survey lands, the entry is scored on the public ledger — hits and misses alike. That is the whole trust model.
Pains, jobs, objections and language — long-form transcripts you can interrogate, from segments too expensive to recruit live.
Kill the weak assumptions before the deck ships. Directional reads on demand, framed as falsifiable claims with segment splits.
Reactions, barriers and willingness-to-pay bands across the panel — treated as directional by design, with the caveats printed on the report.
A 600-respondent study returns in minutes — run it before every decision, then spend live-research budget only on what survives.
Packs launch one at a time — seed, calibration, language and fidelity profile are built per market from that market's own statistics office and benchmark survey. Nothing is inherited from the last one, and nothing is claimed before it is scored.
Interviews in formal Filipino. Seeded from the national labor-force and family-income surveys, raked to census margins across 18 regions; anchored to the central bank's quarterly Consumer Expectations Survey with a decade of backfill.
Seeded from Census Bureau microdata — 100,000 personas with household structure, income, party lean and urbanicity. Anchored monthly to the Michigan Surveys of Consumers and the NY Fed's SCE, with party, region and income crosstabs.
Interviews in usted-register Mexican Spanish. Seeded from INEGI's household-income and labor surveys on CONAPO population projections; anchored monthly to ENCO — the portfolio's best anchor, with person-level microdata released the same morning as the headline index.
Including us. Here is how the category typically answers — and how we do.
No — it's the instrument you run before and between fieldwork. Kill the weak concepts in minutes, then spend your live budget where a real sample is worth $40–80 a complete. When the panel disagrees with reality, reality wins — that's what the calibration and the public ledger are for.
A panel is not a model. Each respondent is a fixed person drawn from official microdata — income, region, household, values — and the panel's aggregate answers pass through corrections fitted to the market's benchmark survey and validated on held-out items. In our own bench of 21 frontier models, none matched the calibrated panel on its anchor.
Before an anchor release, the panel's forecast is frozen and its SHA-256 is published — distributions locked and unpublishable until scored. When the survey lands, the entry is graded and the full artifact opens for audit. Misses stay on the ledger. Retrodiction can't do that; a sealed forecast can't cheat.
Real things, documented. Synthetic panels over-smooth polarization — partisan gaps compress relative to reality — and read optimistic on some sentiment scales; the fidelity profile carries those limits per market, and willingness-to-pay is treated as directional by design. If a vendor tells you their panel has no failure modes, ask who scored it.
The first packs are onboarding a small set of design partners. Tell us what you want to test and where — a human reads every request and replies to your email.
We read every request. You'll hear back at the email you left; no sequences, no drip.