How firm is this year's production figure?

1. How far the first published figure moves

Each edition prints two years: last year as an estimate, and the year before it, revised. No series in this store is printed three times — not one of the 2,254 country series, and none of the world totals — so where a year is measurable, the revision is that estimate against its single revision, as a percentage of the estimate, and every measurable year carries the same one chance to move. 1,763 series (commodity, country or world total, measure, year) are measurable; 746 are not: 507 appear in only one edition (the newest data year, and years either side of a gap in the editions we hold) and 239 lost a printing to a dash, a "W" for withheld or an "NA". Those do not enter any median. The measure, the thresholds and the direction rule were filed before any revision was computed; the last-ten-years column was added afterwards and is logged there as a dated deviation, as is the fact that the filing's “settling” check cannot be computed at all, there being no third printing.

Source: USGS Mineral Commodity Summaries, 1996–2026, read from the annual chapters. Computed values: out/usgs_revisions.json.

Each median is over that commodity's measurable world-total years, counted in the second column. “Largest world revision” is the biggest single move of the printed world total, not of a country. The up / down / unchanged column is the share of revisions in each direction across every mine series in that commodity, countries and world total together — a wider denominator than the other columns, as the filing defines it, and each small country counts the same as the world total. A reprint that does not move counts as unchanged, not as a revision down.

Copper's figure has moved least; antimony's most. Copper's world total ends up 1.5% from its estimate over the whole period and 2.1% over the last ten years — firm on the filed reading, soft on the recent one, which is how much a label can depend on a window. Antimony's moves 11.6%, and once moved 42.5%. Graphite gets worse in the recent decade, not better: 6.9% over the whole period, 8.3% over the last ten.

There is no direction to correct for. The filing licenses one reading, “revised up more often than down” at 60%, and no commodity reaches it. Nor is the rest downward: a reprint that does not move is not a revision down, and unchanged reprints are common. Graphite's mine series are 26% up, 36% down and 39% unchanged; rare earths are 37% unchanged. An earlier version of this page read those as a downward bias, which is wrong on both counts — the filing never defined a downward reading, and the unchanged share was being counted as down.

China's rare-earth figure is the one that moves least: its median revision is 0.4%, so half its measurable years were reprinted with a move smaller than that. That is consistent with a quota figure being copied across editions, and equally with a rounded official number being reprinted; this design cannot tell those apart. That median was 0.0% until the 2004 edition of the chapter was recovered — see the note on retrieval below, and the dated entry in the filing.

The largest single revisions

Source: USGS Mineral Commodity Summaries, 1996–2026, read from the annual chapters. Computed values: out/usgs_revisions.json.

This table ranks by percentage of a small first figure, so it lists small series rather than the changes that move a world total: Kyrgyzstan's antimony is printed as 20 tonnes in the 2025 edition and 700 in the 2026 one, both marked estimates. The levels are shown so the size of each series is visible beside its percentage. The chapters attribute such changes to new reporting — the 2026 antimony chapter says its 2024 revisions for China, Iran, Kazakhstan, Kyrgyzstan and Russia rest on company, Government or third-party reports — which is re-estimation, not necessarily a first measurement. The largest move of a world total is elsewhere: tungsten for 1995, 20,000 to 31,000 tonnes between the 1996 and 1997 editions, with China 10,000 to 21,000 in the same pair.

The other nineteen commodities, added after the fact

The six above were the commodities the panel held when the study was filed, which is not a claim about which commodities revise most. The panel has since grown to twenty-five, so the same measure was run on the rest, under an amendment filed before those numbers were computed — which fixed the floor at 5 measurable years and committed in advance to reporting the ranks either way. The endpoints hold: Antimony still moves most and Copper least, and no added commodity falls outside that range. The six do not sit together at the extremes, though. Their ranks of twenty-five are antimony 1, graphite 5, cobalt 10, tungsten 11, rare earths 21, copper 25, and four of the six largest are commodities the amendment added (ranks 2, 3, 4, 6) — so the added set does not simply fill the middle.

Source: USGS Mineral Commodity Summaries, 1996–2026, read from the annual chapters. Computed values: out/usgs_revisions.json.

Three things in this table are not what the ranking makes them look like, and all three were caught by the reviewers rather than by the guard.

First, the ranks compare different stages. Five of the added chapters print no mine table at all: the USGS tabulates germanium, indium and tellurium at the refinery stage, and gallium and magnesium as a single production line. Each commodity is therefore read at the stage its own chapter prints, named in the second column — but a refinery revision and a mine revision are not the same quantity, so a single list of twenty-five is a list of "each chapter's primary-stage world total", not of comparable things. The twenty mine chapters ranked on their own: antimony 11.6%, lithium 7.3%, chromium 6.9%, graphite 6.9%, fluorspar 6.6%, feldspar 6.5%, vanadium 6.0%, cobalt 5.7%, tungsten 5.3%, barite 4.9%, tin 4.5%, titanium 4.5%, manganese 4.2%, molybdenum 3.6%, zinc 3.0%, rare earths 2.8%, lead 2.6%, phosphate_rock 2.3%, nickel 2.1%, copper 1.5%. On that same-stage reading the filed six still hold both ends, and the commodity worth naming is lithium, second at 7.3% on 25 measurable years — the same window as copper, and above three of the six that were filed. Gallium, ranked 2, and Tellurium, ranked 8, are not in that list at all — they are refinery or single-production-line chapters, and none of them belongs in an order with mine figures.

Second, copper keeps its place on the recent window but loses its label. Copper is still the lowest of the twenty-five over the last ten printed years, at 2.1% — but that is above the 2% cut, so on that window it is soft, not firm, and the next commodity, copper at 2.1%, is a tenth of a point behind it. The amendment said that if copper stopped being the firmest the page would say so: it has not stopped being the lowest, and it has stopped being firm. Both bands are printed for every commodity so that the difference is visible rather than argued.

And the direction rule bites once the set is widened, on the filing's own wording. The filing reads the direction as "the share of revisions that are upward", at 60%. The code had been dividing by every printing instead, unchanged reprints included, which is a looser test; on the filing's denominator three commodities clear it, both of them added: gallium 86% of its revisions upward (7 measurable years), indium 66% of its revisions upward (24 measurable years), chromium 61% of its revisions upward (26 measurable years). Neither is a commodity to build an argument on — gallium's is seven years of a production line, and indium's twelve of a refinery series — but the reading is licensed by the filing and is reported. None of the six filed commodities reaches 60% on either denominator (the highest is rare earths at 59% of its revisions), so section 1 stands as published.

The “last 10” window is the ten years to each commodity's own last measurable year, and the dates printed beside it are the first and last measurable year inside that window — so a span shorter than ten years means years inside the window that could not be measured, not a different rule. It is 2015–2024 for eleven of the twenty-five; the four exceptions are phosphate_rock 2008–2017, germanium 2011–2020, tellurium 2017–2024, gallium 2018–2024. Germanium is the one whose window ends early, and for two reasons that are both in the source: the 2023 edition prints its country table with every cell "NA" or "W", so 2021 and 2022 have no usable second printing, and from the 2024 edition the chapter prints no world germanium figure at all — not a table, and not a number in the text. Gallium is thinner still and for a third reason: only the editions from 2019 on print a country table, the 2010–2018 chapters give a world estimate in a sentence, and the chapters before that say the data are unavailable. Seven measurable years is what a chapter that began tabulating countries in 2019 can support. And gallium's table changes its own caption inside the comparison: through the 2024 edition it is "World Production and Reserves", and from the 2025 edition "World Low-Purity Production and Production Capacity". The 2023 pair straddles that change, and gallium's largest revision — 11.6% for 2024, 760,000 to 848,000 kg — falls entirely inside the new caption. This study checks that a unit change is not read as a revision; it does not check for a caption change, so gallium's rank should be read as provisional on that ground alone. Its largest revision is also its newest measurable year, which is the one case here where the record does speak to a current figure.

A retrieval note, because it changed a published figure. Three of these counts moved while this section was being reviewed, not because the USGS revised anything but because the atlas had failed to fetch chapters that exist. The germanium 2019 chapter is published as mcs-2019-germa_0.pdf, and the fetcher's pattern did not allow the trailing _0; eleven indium chapters and the 2004 rare-earth chapter are linked from their pages with a duplicated path segment that returns 403, and were never retried under the path the USGS actually serves; and the germanium, tellurium and indium chapters for 2003 sit in the yearly volume and had not been cut out of it. All are now in the panel: germanium moves from 16 measurable years to 19, indium from 12 to 24, tellurium to 14 — and rare earths, one of the six filed commodities, moves from a published 2.5% over 24 years to 2.8% over 26, with its China median going from 0.0% to 0.4%. Its band does not change. That correction is logged in the filing, and it is the third time in this study that a number was wrong because of what had not been fetched rather than what had been computed.

Thin coverage, stated rather than smoothed over: gallium 7, tellurium 14 measurable years, against twenty-five for copper. A median over seven years and a median over twenty-five are not the same quantity, and every rank here should be read with the year count beside it. The up / down / unchanged shares are over every series in the commodity at its primary stage, countries and world total together — a wider denominator than the medians, which are the world total alone — and they are rounded, so a row can sum to 99 or 101. Germanium shows how far those two denominators can sit apart: 52% of its series are reprinted unchanged while its world total still moves a median 3.0%, and its largest single world revision is 74%. Lithium's largest is 60.5%. Neither is a bound on this year's figure, for the same reason as in section 1.

One table had to be repaired to be read at all: magnesium's "World Primary Production and Reserves" is parsed as production through the 2020 edition and as smelter from 2021, because the printed row label changed while the table did not. Read as two measures it became two half-series that never meet. They are merged only where the evidence says it is one table — the same printed caption, and no edition printing both — which copper's mine and refinery tables fail, six editions printing both. The one year measured across the rename is 2019, and it moves 1.8%, so the merge is not manufacturing a revision out of a renamed row: dropping that year moves magnesium's median by less than a hundredth of a point, from 4.53% to 4.54%. A caption can stay while a definition changes, so this is evidence, not proof.

2. Mining and refining are different maps

The USGS prints a refinery table by country for copper only. It is worth reading on its own — bearing in mind that 2025 is a first printing, the kind of figure section 1 is about.

Source: USGS Mineral Commodity Summaries, 1996–2026, read from the annual chapters. Computed values: out/usgs_revisions.json.

In the 2025 table China accounts for 8% of reported world mine production and 48% of reported refinery production. Chile is the mirror image: nearly a quarter of world mining, a small share of refining. Japan and Germany appear in the refinery column with a dash in the mine column, not a zero; Peru and Zambia mine copper they barely refine. A dependence read off mining alone misses where the smelting and refining sit — the argument this atlas is built on, here from a single table rather than a chain of inferences.

What these two columns are not: refinery output is where metal is produced, not who owns it, so a Chinese-owned mine abroad counts in its host country's mine column. Refined output includes metal made from scrap, which no mine column can show. Cathode won by leaching at the mine is counted in both columns, so the two are not disjoint. Chile's gap is concentrate shipped out to be smelted and refined elsewhere. And neither column says anything about semi-fabrication, where another map again applies.

3. Where the two agencies disagree

The US Geological Survey and the British Geological Survey both publish world production. The gap below is the median, over the years both cover, of |USGS world total − BGS world total| divided by the BGS total, using BGS production rows only. It is unsigned and says nothing about which agency is right. Neither series is a correction of the other, and a gap can be measurement, definition, coverage or vintage; the number of countries reporting each BGS series is shown, because a series carried by few reporters is a different animal from one carried by many.

Source: as above, against BGS World Mineral Statistics as held in the atlas. Computed values: out/usgs_revisions.json.

Copper is the close case: 0.6% apart on mining over 2002–2024, and 1.2% on refining — though the refining comparison rests on 2019–2024 alone, because the USGS chapter has printed a refinery world total only since then. Graphite is the far one at 30.7%, and not because the baskets differ: both series are natural graphite and both exclude synthetic material. China accounts for much of the early distance — in 2010 BGS put Chinese output at 1,800,000 tonnes against the USGS's 600,000 — and that difference has closed: by 2024 both print 1,270,000 tonnes. But the world totals still sit 10% apart in that year, so the rest is other countries, coverage or rounding, and 30.7% is the median over 2002–2024, not today's gap.

Rare earths compare at 5.7%, but on a BGS series carried by a median of 7 reporting countries against a USGS table that carries 3–13 countries with output over the same years, so coverage rather than measurement may be doing the work. Whether both sides use the same contained-metal definition is not checked here for cobalt or antimony.

What we do with this

Three rules, adopted with this page and applied from here on rather than claimed of what is already published. A current-year figure is quoted with both revision medians beside it, the whole period and the last ten years, and never with the band presented as its precision: firm, soft and weak are labels for a past record, not error bars. A claim that turns on a change about the size of that record is checked against the revision history before it is made rather than taken at face value — for antimony that means a change of a few per cent is no larger than what the source has moved on its own. And when the other agency is cited, the page names it and names the years.

What this cannot say: it measures how the published figure moved, not how close either version is to the truth — both could be wrong in the same direction. Revisions within a year are not independent across countries, since the world total is revised with its parts, and successive years share a method, so they are not independent draws either. A country that falls into "Other countries" in a later edition drops out of the count for that year. The medians cover twenty-five commodities over unequal windows — tungsten 29 years, gallium 7 — and the copper median in section 1 is mine production, not the refinery series section 2 reads. The agency gaps are differences, not a reconciliation: whether both sides use the same contained-metal or concentrate basis is unchecked for every commodity here.

What we did with this measurement. It was turned on the atlas's own work: the concentration study's claims were recomputed with these revisions resampled into the production behind them, and one of them turns out to sit close to the noise. Auditing our own claims.

Method and data

1. The figures. US Geological Survey, Mineral Commodity Summaries, editions 1996–2026, the annual chapter for each commodity (a US government work). usgs.gov. Read into a panel by pipeline/parse_usgs_mcs.py, which reads the printed page geometrically: a footnote marker is set against its value, so in plain text "7100,000" cannot be told from a number, while on the page the marker is smaller type.
2. The comparison. British Geological Survey, World Mineral Statistics, as held in the atlas. bgs.ac.uk.
3. The filing. The measure, the thresholds and the direction rule, committed before any revision was computed, then changed in fifteen dated deviations — among them the production-only filter on the BGS comparison, a graphite world total that had been multiplied by 1,000, the rare-earth series and unit line, and the separation of unchanged reprints from revisions down. One discarded row was wrong by three orders of magnitude; the median contained it rather than ignoring it — the graphite gap would have printed 31.1% instead of 30.7% — which is why the log, not a calm median, is the check: usgs-revisions/PREREGISTRATION.md.

Built by build_revisions.py from out/usgs_revisions.json, written by usgs-revisions/analysis.py from pipeline/data/usgs_mcs_history.parquet.