Not a promise — a build property
Ask a substantive question. Get a statistically correct, permanently citable answer — computed from governed public data, not recalled from a model's memory.
In 2024, said abortion should be legal for any reason — up from 39.5% in 2006.
Every number on the platform resolves to a receipt like this one. Click any statistic, see where it came from.
66 essays published so far — the same discipline behind every one.
Library-vendor AI assistants are literature tools — they discover and summarize documents. None of them runs a weighted estimate.
General-purpose AI assistants answer statistics questions fluently, and fabricate freely — wrong numbers, wrong qualifiers, wrong vintages, stated with total confidence.
Librarians are professionally accountable for source reliability, and are being handed AI they can neither audit nor disable.
“Ask a question, get a design-correct estimate with permanent provenance” is an empty category. We built it.
Four steps, the same for every published figure.
A frozen, immutable copy of the source data. Every computation pins to one version, not a live query that can drift under you.
A versioned, design-correct computation — the right survey weights, the right strata, suppression rules applied before anything is shown.
The number, with its full provenance, at a permanent address. A citation made today still resolves a decade from now.
One click to APA or BibTeX. A student, a journalist, and an AI answer engine all cite the same underlying object.
Before anything is computed at scale from a source, the full analysis plan — weights, demographic cuts, suppression floors, where a series breaks — is rendered into a review packet. A human statistician reads it and signs. The approval is tied to a fingerprint of the exact plan; edit the plan and the approval voids, reopening review. The build software refuses to run an unapproved plan.
One design mistake would replicate across thousands of published numbers. So a person signs every plan before a single number is computed.
An AI tool a library can put its name behind.
A Black president and a killing in Minneapolis were supposed to bend the arc of American racial opinion. The survey record says one barely registered — and the other moved the country sharply, briefly, and only on one side of the aisle.
GSS + ANES (1972–2024)
CES (2008–2024)
Precinct returns · 2016–2024
Between 2008 and 2019 — before anyone had heard of COVID-19 — the share of humanity reporting a stressful day rose by ten percentage points. Enjoyment never moved.
Gallup World Poll (2008–2019)
GSS (1972–2024)
Gallup Daily Poll (2009–2017)
You would expect the map of American disability to be a map of where old people live. It is not — disability nearly doubles from the most to the least advantaged neighborhoods, while the elderly are spread evenly across both.
CDC PLACES 2024
CDC PLACES 2024
CDC WONDER · Linked Birth/Infant Death (2018)
Across 147 countries, money reliably buys the feeling of financial sufficiency — and the feeling never tops out. But whole regions feel far richer, or poorer, than their income says they should.
Gallup World Poll (2016–2020)
Federal Reserve · Survey of Household Economics (SHED)
In the 1970s American women called themselves happier than men. A famous study watched that lead vanish. Extend the data eighteen more years and the decline doesn't continue — it dissolves into noise.
GSS (1972–2024)
Gallup (2.46M interviews)
Gallup World Poll · 143 countries
Interactive tools for the underlying data — drill into demographic breakdowns, filter by year or region, compare across institutions.
Every essay publishes as structured Markdown and machine-readable JSON-LD. Your agents can fetch methodology, read the claims, and source the data directly.
