What makes a figure trustworthy?
Not that an authority published it. Authorities publish figures that disagree with each other constantly, and the same estimate is republished under several logos until it looks like agreement. A figure is trustworthy when you can see where it came from, what it counts, who else says otherwise, and what was never checked at all.
That is the whole job of this engine, and every claim on this page is one you can test against the live service rather than take from us.
Mixing sources without laundering them
The hard part of combining datasets is not fetching them. It is knowing when they may be put in the same sentence.
One explicit door for non-UN data
A figure from outside the UN system enters through a single argument,
external_variable_dcid, and is labelled EXTERNAL
everywhere it appears. It faces the same comparability checks and the same
lineage resolution as a UN figure.
It can corroborate a UN figure. It can never replace one, and it is never quietly folded into a total.
The aggregator is not the source
An external figure is traced through to the body that actually produced it. Ask this engine for Germany's population from outside the UN and the record names Eurostat, not the aggregator that served it.
Citing the middleman is how a national statistic loses its author.
Thirteen ways two figures can fail to be comparable
Before any two numbers are compared they are checked on unit family, scale, denominator, population constraints, observation period, reference date, geography, measurement basis, release vintage, coverage, currency basis and index base.
Nine of those withhold the verdict when undeclared. Four are disclosed on the record instead, because a warning raised on everything is a warning nobody reads.
Provenance you can recompute, not provenance you are asked to believe
Every record carries the chain that produced it, and the instructions to rebuild it without us.
Publisher count is not origin count
Three agencies publishing one inter-agency estimate are one source. Every record states both numbers on its face, and names the institution the agencies resolve to.
Where an origin cannot be established, the record says unresolved rather than guessing. Unresolved is not the same as independent.
A replay manifest, not a promise
Each sealed record lists the exact upstream calls that produced it, against a named endpoint and a pinned protocol version. You can run them yourself and compare.
The rules that produced the verdict are hash-pinned, and the record cites their SHA-256, so you can tell which rules applied.
What was never checked is written down
This build does not assess publication recency, because the graph carries no publication or revision date, and it does not assess uncertainty intervals. Both are stated on every record rather than quietly passed.
Because uncertainty is unassessed, citation readiness never reaches
READY_FOR_CITATION on its own. A human signs that off, not us.
Transparency and repeatability
The engine contains no language model. The same question and the same rules produce the same answer, every time, for anyone.
Hash-chained, so tampering shows
Each step of the reasoning is SHA-256 linked to the one before it. Editing any step breaks every hash after it, and the break is visible to the reader rather than to us.
Verified in your browser, not on our server
The verifier recomputes the whole chain client-side and sends the record nowhere. Routing that through our server would mean asking you to trust our server to tell you whether to trust our server.
Write-once records outlive the data
A sealed record is immutable. If a publisher revises a series next year, the passport still states what was served on the day, so a revision becomes a visible difference instead of a silent edit.
FAIR, self-assessed and falsifiable
Read live from the engine at /fair, so this page cannot
drift from what the service actually reports. Where this build falls short it
says so, and says what is missing.
From a number to a decision
We stop one step short of deciding for you, on purpose.
The distance between a harmonised database and an informed decision is the part everyone finds hard, and it is also exactly where a language model will invent something plausible. So this engine does not rank sources, does not average them, and does not recommend an action.
What it does instead is close the part that can be closed honestly. Given a verdict, a lineage and a set of failed checks, it states the sentence you are licensed to write, and the one you are not: that you may cite this figure attributed to UNICEF, and may not call it “the UN figure”; that you may report both values and say they disagree, and may not average them.
That is a smaller claim than a recommendation engine and a much more useful one, because a decision-maker who knows precisely what their evidence supports can make the call themselves, and defend it afterwards. See it on a real question →