How AgricultureID handles conflicting official sources

Two official sources disagree. One is a national statistics agency, one is an intergovernmental body republishing that agency's data. The numbers differ by four percent. Which one goes on the page?

The tempting answer is "the more authoritative one", and it is wrong twice over.

Why silent resolution is a fabrication

Choosing between two sourced values produces a third thing: a judgement about which source is better for this quantity, in this country, in this year. No source made that judgement. If the page shows one number and cites one source, the reader is told there is one answer, and the existence of the other is destroyed.

The reader can no longer tell that this figure is contested. That is information, and it was in the corpus a moment ago.

Most conflicts are definitional

When we look at actual disagreements, very few are one source being wrong. Nearly all are two sources measuring different things under the same word.

Land "equipped for irrigation" and land "actually irrigated" are both irrigation figures for the same country in the same year, and they are not the same number — one is infrastructure and one is use. A trade flow recorded by the exporter and by the importer differ because they are measured at different borders, at different times, with different valuations. A national average cost figure and a regional one differ because they cover different farms.

Resolving these by picking a winner does not produce a truer number. It produces a number whose definition nobody can recover.

What we do

Record both, with what each is. Every value carries its source, its capture, and the source's own words for what it measures. Where two are held, both are held.

Say which is displayed and why. Where a page must show one figure, the choice is a stated rule — the reporting country's own statistic over a republication, for example — not an unstated preference.

Check the source against itself first. Many apparent conflicts are ingest errors. Before recording a disagreement between two sources we verify each source's internal arithmetic: that a published total equals the published components, that categorical shares sum to a whole, that a release identifier is the one the payload carries. A surprising number of "conflicts" disappear here.

Publish nothing as certain where the conflict is unresolved. The platform has a vocabulary for this, and pages use it: verified, partial, and several distinct kinds of "not yet indexed". A conflict that cannot be resolved from evidence is not rounded down to silence and not rounded up to a claim.

The case that made this concrete

FAOSTAT publishes an area called "China" alongside China mainland, Hong Kong SAR, Macao SAR and Taiwan Province of China — all reported separately. Its value equals the sum of the four in every one of the 25 series where they coexist.

Read naively, the corpus would have held both the parts and the whole in one country column, and any total across countries would have counted the same land twice. There was no conflict of fact here at all. There was a conflict between what the column looked like — a country — and what one of its rows meant.

That is what most source conflicts turn out to be, and it is why we resolve them by reading the source more carefully rather than by ranking sources against each other.