Everything here is built from public data by script, and rebuilt from scratch each time. There is no hand-entered spreadsheet anywhere in the chain: if a number appears on this site, a named source and a named calculation produced it, and both are written down below.
The engine, step by step
Five operations turn raw public files into the figures on the site. Each is stated with its actual rule, because “we reconcile the data” is not a method.
Decide who to believe
Every shipment is reported twice — once by the exporter, once by the importer — and the two rarely agree. Import values also include freight and insurance that export values do not, so they are not even measuring the same thing. The engine first removes that wedge by estimating the CIF/FOB markup, then combines the two sides. Version 1 uses an equal-weighted geometric mean of the adjusted pair.
Reporters are also scored, so that combination can improve. Each country’s systematic bias is estimated by robust two-way median centering (removing partner and product effects), and its reliability = 1 / residual variance, shrunk toward the average so a country with few flows cannot look artificially precise.
flows_reconciled — one agreed value per flow, plus a quality score per reporterCount each shipment once
Seven sources overlap. The same tonne can appear in a national release, a mirror record and a compilation. Records are matched on flow identity and de-duplicated to a single best row per flow, with the losing rows kept rather than deleted so the choice can be inspected.
flows_bestApply the physical limit
Trade data says where a shipment came from, which is not the same as where the material was produced — a hub can export what it never mined. So the atlas imposes a physical constraint: a country cannot ship out more than it produced and did not itself consume.
The weight w is how much we trust the trade record against the physical prior. It is set at 0.5 and documented as a knob, not a discovery — which means a corrected share is the midpoint of a range, never a point finding, and the pages say so.
Estimate what a country actually uses
Consumption is mostly unpublished, so it is estimated from what a country visibly does — how much steel it makes, how many cars, how much construction — times how much of each material that activity needs. The intensities are not guessed. Each is back-solved from a known world total, so the country cells can be uncertain but the world sum cannot drift:
This is deliberately trade-independent: it never looks at imports, which is what lets it be used as an independent check on the trade data in step 3 rather than a restatement of it.
Turn it into indices
Shares become comparable measures: concentration (HHI of national shares), a weighted supply-risk score, resilience under supplier removal. Where an input is an interval rather than a measurement, the index carries a score band and a rank band instead of pretending to a point.
What each analysis takes in, and puts out
Six analyses, each attacking supply risk from a different side. Same question for each: what goes in, what is done to it, what comes out.
1 · Value chains
2 · Chokepoint map
3 · Risk & scenarios
4 · Trade engine + anchor
5 · Concentration
6 · Cross-source comparison
The five checks a number passes before it appears
This is the part worth knowing before you use anything here. Most publishing on critical minerals gives you a number. The atlas gives you the number and what it is worth — and takes numbers down in public when they fail. Every figure runs this gauntlet.
Nothing is published if a link is broken, a dataset a page depends on is missing, or a retracted figure has reappeared. This runs before every single deployment and has never been bypassed.
Every estimate carries a clickable basis note naming the source, the quoted row and the computation. Reported, computed and estimated are never silently mixed — and a figure with no recorded source says so on its face rather than borrowing the authority of the ones that have it.
Findings are attacked before publication, not after. The concentration finding went through five review passes and two complete framings were discarded, because a claim that only survives friendly reading is not a finding.
A corrected source does not correct its copies. Every derived figure is compared against the source it was copied from, so a page cannot keep publishing a number the rest of the site has already retracted.
A sum over a handful of reporting countries is not a world total, however plausible it looks. 13 of 53 materials fail that test — including lithium and rare earths — and the pages that use them say so rather than quietly rounding the limitation away.