How the record was built

What counts as an entry

An entry is a specific thing people did for the first time or did notably better: an invention, a discovery, a world first, a major improvement, a cultural milestone, a social or legal innovation, or a feat such as a voyage or an ascent. Each entry is credited to the present-day country where the work was done, with any other countries that share credit listed alongside. Multinational projects such as CERN are credited to International.

How it was compiled

The record was compiled in October 2026 with an AI model, Anthropic's Claude, working country by country and period by period from its training knowledge. Each section was written to a common specification, checked for format, and then merged. Near-duplicates (the same event described twice under different names) were found by comparing names, descriptions, years and countries, and merged into one entry. In this second edition another 713 duplicates were found and merged, as described below. The result is 38,331 entries from 197 countries and territories.

Because the entries come from a language model's knowledge rather than from a team of historians working through archives, treat any single entry the way you would treat a well-read reference book: usually right, sometimes wrong in the details, and worth checking against its sources before you rely on it.

How government types were assigned

Each entry is classed by the government in power where and when the work was done. The class comes from a timeline of each country's main regime periods, refined by the historical state named in the entry, so Soviet work counts as communist and Venice or Florence count as republics. There are eight classes:

Democracy

Competitive elections with broad suffrage choose the government. Includes parliamentary monarchies such as today's Britain or Japan, and classical Athens.

Oligarchy or limited government

Elected bodies hold real power but a narrow elite or electorate chooses them: the Roman Republic, Venice, the Dutch Republic, Britain from 1689 to 1917, the early United States.

Monarchy or empire

Hereditary or self-proclaimed rulers with no effective elected check.

Authoritarian regime

Military, fascist, personalist, theocratic or non-communist single-party rule, and regimes whose elections aren't competitive.

Communist state

Marxist-Leninist one-party states such as the Soviet Union and the People's Republic of China.

Colony or dependent territory

Ruled by a foreign power after 1450, including protectorates, mandates and wartime occupation.

Tribal or pre-state society

Before the first state in the territory, or where the form of government is unknown.

International

Multinational projects.

On the home page and in the combined view, the first two are grouped as democratic and the next three as autocratic. Borderline cases follow simple rules: "democracy" requires competitive elections with at least near-universal male suffrage, so Britain counts from 1918, the United States from 1828 and France's Third Republic from 1870. The classes are deliberately broad, so treat the shares as approximate.

How the sources were found and checked

Every entry was given up to four suggested sources, preferring at least one primary source (the original paper, patent, official record or text) plus scholarly secondary works such as university-press books and peer-reviewed articles. General encyclopedias are allowed as one source at most and never alone. Wikipedia and sites that copy it are never used. Where no source could be named with confidence, the entry says so instead of guessing.

• 83,116 suggested citations in all.
• 31,541 entries have two or more suggested sources, 3,898 have one, and 2,892 have none the model was confident of. These counts measure suggestions, not verified sources.
• 22,660 entries carry a note on disputes over priority, dating or credit, and 251 a note where the sources disagree with the entry's date, place or people. Both are shown on each entry.

The citations were also written by the AI model, from its knowledge, without web access. To measure how reliable they are, 380 citations were drawn at random and checked by a separate pass with web access against library catalogues, publisher pages, DOI records and patent offices:

• 338 were confirmed as real works that cover the entry.
• 8 were real and relevant but had a detail wrong, such as a year or page range, and were corrected.
• 25 were confirmed to exist, but whether they cover the specific entry couldn't be confirmed.
• 9 were wrong, mostly real works about a neighbouring event, and were removed or replaced.
• None were invented works.

Each citation that was checked is marked in the downloads and on the Browse page. The rest haven't been checked one by one, so the sample rates above are the best guide to how far to trust them. No entry has yet been checked by a person.

The automated second pass

A separate automated research pass then went through every entry. It searched Wikidata, the reference lists of matching Wikipedia articles and OpenAlex, an index of scholarly papers, fetched what it found, and recorded whether each page or abstract mentions the entry's key terms. It also flagged entries whose sources seemed to give another year, whose government type looked wrong, whose claim to be first seemed questioned, or which looked like duplicates. It ran twice: a first run over every entry, and a second, longer run in October 2026 over the entries still without two independent sources. Every flag from the first run was reviewed; the second run's much larger set was screened and sampled as described below. The decision and reason for each flag is in edition_changes.csv.

• Links. 5,332 links on 4,426 entries are shown as found and matched automatically. From the first run, links were kept when at least two of the entry's key terms matched, and left out when they point to library catalogue, census or search pages, which match on a name rather than the claim. The second run fetched about 28,000 more pages and abstracts, found through the citations of related Wikipedia articles and through keyword searches of Crossref and Europe PMC. Spot checks showed that most of its keyword-search hits share the entry's words without documenting it (6 of 16 readable ones did, even when the year and country also matched), and so do pages that don't confirm the entry's year and country (8 of 21). Pages from Wikipedia's citations that confirm both mostly do (18 of 23), so only those were added: 3,292 links, marked as checked for year and country. A matching page is a lead, not proof.
• Dates. In the first run 206 entries were flagged because a source mentioned other years near the claim. On review none was a different date for the same event: the numbers were page and catalogue numbers, measurements, publication years, the ends of a date range, or the dates of other events. The second run flagged 2,254 more. Most came from pages that don't support the entry, or gave years far from its date. In the 482 cases where a supporting source gave a nearby year instead of the entry's, a random sample of 60 was read and none was a different date for the same event. No years were changed.
• Government types. 675 entries were flagged because the government at the time differed from the start of the country's current constitution, such as France's Third Republic before the 1958 constitution, or Athens before modern Turkey. That comparison isn't a test of the classification, and 674 of the 675 were right. One was corrected.
• Duplicates. The pass found 19 groups of entries with the same name. Thirteen were the same event entered twice and were merged, keeping their sources. The rest were kept as separate entries and renamed: independent discoveries made at the same time in different countries, or an object and its later excavation. Because the pass only caught identical names, a follow-up scan compared similar names across the whole record and found 142 more duplicates, such as separate entries for the start and the completion of the Thames Tunnel; each was checked and merged the same way. A third scan compared the wording of every entry's name and description with entries from the same country and period, which catches the same event described in different words, such as "Champlain founds Quebec" and "Founding of Quebec". Each of the 949 closest pairs was read by hand, and 557 entries that repeated another were merged into it. The rest were kept apart because they record different steps, such as a patent and the first bridge built with it, or the same discovery made independently by separate teams.
• Firsts. 344 entries were flagged because a source used words such as "predates" or "contested". 26 already said the claim is disputed and seven got a new note, such as the Mauch Chunk railway's claim to be America's first roller coaster. In the rest the word was about something else, the match was on a word such as "predator", or the earlier rival doesn't change the claim as worded.

The pass's raw output, including its Wikidata matches, coordinates and images, is in the downloads. Its Wikidata matches aren't used on the site yet: a sample of even its most confident matches included wrong articles.

What the counts can and can't show

The charts on this site count entries. That makes them a measure of this record as much as of history, and several things other than freedom push the counts up or down:

• Documentation. Countries whose history is well written up in English, and recent decades, are easier to cover in depth. The record almost certainly over-represents Europe and North America for that reason, which flatters the democracies of the last two centuries.
• Population and wealth. The counts aren't adjusted for how many people lived under each kind of government, or how rich they were. Richer countries have more scientists, more patents and more museums, whatever their government.
• Most of history was monarchy. Before about 1800 nearly every state was a monarchy or empire, so most early achievements are credited to them. Comparisons are only fair within the same era.
• Strong counterexamples exist. The Soviet space programme, Song dynasty China, the Abbasid golden age and Meiji Japan all produced a great deal under unelected rule. Any argument about freedom and progress has to account for them.

The fairest tests compare like with like: free and unfree countries in the same decades, measured per person, or natural experiments such as East and West Germany or North and South Korea. Scholars disagree about how much institutions explain. Joel Mokyr and Daron Acemoglu and James Robinson argue that open, inclusive institutions drive sustained innovation; others point to state-led successes as evidence that the picture is more mixed.

Corrections

If an entry is wrong, missing, credited to the wrong country or government, or cites a source that doesn't support it, email [email protected] with the entry's name and what the sources say. Corrections are folded into the next edition.