Per ora questa pagina è in inglese.
How it works · rules version 0.4 (14 Nov 2026)
Every number has a rule, and every rule is here
No score on this site comes from an AI's opinion. Scores come from published data and the fixed rules below, applied the same way to every country.
Versions of the rules
Each version of the scoring rules has its own page and a citable PDF that never changes, with the data, the formulas, a worked example, the checks and the limits.
- Scoring rules 0.4: method paper · PDF · released 2026-11-14 · in force
- Scoring rules 0.3: method paper · PDF · released 2026-11-14
- Scoring rules 0.2: method paper · PDF · released 2026-11-01
- Scoring rules 0.1: method paper · PDF · released 2026-10-12
Three voices
Say — what governments say about each other in their General Debate speeches at the UN, since 1946. We show the sentences themselves.
Do — what governments do: how they vote at the UN, how much they trade, whether they fought on opposite sides or the same side, which clubs they share.
Live — how people are connected: migration, refugees, and football. This layer is shown but never scored: people moving between two countries says nothing about whether their governments get along.
The four scores
UN votes (−100 to +100)
For the last three years with data (2022–2024), we take the two countries' vote agreement on recorded General Assembly votes on whole texts: the same vote counts 1, an abstention against a yes or a no counts ½, a yes against a no counts 0 (the agreement score of Voeten et al., Lijphart's index). Until 8 November 2026 this page called it "the share of votes on which both cast the same vote", which overstated how strict it is; the numbers were always computed this way. Then we ask: among all the countries A votes with, where does B stand? And where does A stand among B's partners? The score is the average of the two positions: +100 means each is the other's closest voting partner, −100 the most distant. We compare with each country's own partners because most countries vote alike on most resolutions, so a raw percentage hides the differences that matter.
Trade (0 to 100)
Average yearly trade in goods for 2023–2025, as a share of each country's total trade. We use the share of the more dependent side, on a logarithmic scale: 0.1% or less scores 0, 25% or more scores 100. Trade cannot be negative: no trade means no tie, not hostility. Sanctions are not yet included.
Security (−100 to +100)
From the Uppsala Conflict Data Program (data to 2025). Fighting on opposite sides pulls the score down: −100 if both were main parties, −60 if one was a main party and the other sent troops to support a side, −30 if both only supported opposite sides of a third party's war. These count fully for 0–4 years, half for 5–14 years, a fifth for 15–29 years, and not after that. Any recent fighting on opposite sides outweighs any alliance. Otherwise: both in NATO +50, and fighting side by side (as a main party or with troops) +30 on the same time scale, up to +80. Peacekeeping contributors serving in the same mission do not count as allies.
Shared clubs (0 to 100)
+25 for each shared membership: EU, African Union, Arab League, ASEAN, BRICS, G7, G20, Gulf Cooperation Council, the five permanent Security Council members, and the same UN regional group. Up to 100.
The index
A weighted average of the four scores. Default weights: UN votes 35%, Trade 20%, Security 30%, Shared clubs 15%. On every pair page you can change the weights and see the index move. The index is shown only when a pair has both UN vote and trade data; otherwise it says Not enough data.
Labels: +55 and above Very close; +25 to +54 Close; 0 to +24 Neutral; −30 to −1 Distant; below −30 Opposed. Who a country is follows the UN: the government that holds its UN seat. Where UCDP calls another authority the government (Yemen from 2015: the Ansarallah (Houthi) authorities; Afghanistan in 1996–2001 and from 2022: the Taliban), that authority counts as a non-state party, and fighting against it is shown as context, not as a war between two states. Since rules 0.4 a pair in which one government fought such an authority gets no index and is labelled Not computable: who the relationship is with, and which rules apply, is unclear there. One exception for labels since rules 0.3: when the security score is −60 or lower, meaning the two are recorded on opposite sides of armed conflict in the last five data years, the label is Distant at most, whatever the index. These describe the data, not the countries or their people.
How we use AI
A language model runs on our own server (currently qwen3.8:27b, through Ollama); nothing you type and none of our data is sent to an outside AI service.
- Ask — the AI answers a visitor's question about one pair, using only that pair's data, and names the section each claim comes from. Every number in its answer is checked against that data; a number that does not match is flagged in plain sight.
- Smart search — understands questions like “How do Iran and Saudi Arabia get along?” and opens the right pair.
The “In short” summary at the top of each pair page is not written by AI. We tested AI summaries and the local model sometimes reversed facts (for example, which side depends on the other), so the summary is written by fixed rules from the same numbers. AI never sets or changes a score, never rates a country, and its answers are always labelled. When in doubt, the data on the page is the source of truth.
Why words are not scored yet
Nations on Record labels each sentence in which one country names another (accused of aggression, criticised, shown as victim, supported, neutral). Most labels come from a language model and are still being checked by people. Until a pair's sentences are checked, we show the sentences and keyword counts, but we do not turn them into a score. A keyword in the same sentence (for example "aggression") says nothing about who is accused of what.
How sure we are
Every pair page shows, next to each score, how much data stands behind it: High, Medium, Low or None. The level never changes a score; it tells you how far to trust it. The rules are fixed and the same for every pair.
- UN votes: the number of shared roll-call votes on whole texts in the years the headline share covers (the last three years with data). 150 or more is High, 50 to 149 Medium, fewer Low, none None. For a share near 50%, that is a sampling range of about ±8 points at 150 votes and ±14 at 50. The typical pair has about 240.
- Trade: how many of the three scored years report goods trade between the two: three High, two Medium, one or none Low. No trade series for one of the two means None. Goods only, from customs data; the data cannot tell zero trade from trade that was not reported.
- Security: High for every pair, because the Uppsala Conflict Data Program covers every state; it counts only conflicts with at least 25 battle deaths a year.
- Shared clubs: High; the list of groups is fixed and checked by hand.
- Index: as sure as the weaker of votes and trade, the two parts it requires.
- Topic bars show a thin line: the 95% sampling range (Wilson) given the number of votes. Votes are not independent draws, so treat it as a lower bound on the uncertainty.
- Voting positions: from the ideal points of Bailey, Strezhnev and Voeten, each with a 90% range. If the two ranges do not overlap the positions are clearly different; if they overlap, the votes alone cannot tell them apart.
Other rules
Eras on the timeline are found in the voting series itself: we split it where the average changes most (binary segmentation, at least six years per era, at most four eras).
Quotes in "In their own words": for each decade, the first sentence in the latest speech of that decade in which one country names the other. Same rule for every pair and both sides. Every other sentence is one click away.
Moments on featured pairs are editorial context with a background link. They never enter a score.
Games never use a pair whose governments fought on opposite sides, as main parties or with troops, in the last five years of UCDP data.
Surprises on the home page are plain rules over the scores, written under each list.
What we can't tell you
- UN votes and speeches are what governments say in one forum. Bilateral diplomacy, embassies and most treaties are not in the data yet.
- The site covers the 193 UN member states and the two non-member observer States, Palestine and the Holy See: the countries the UN itself lists and the data sources follow. Places whose status is disputed, such as Kosovo, Taiwan and Western Sahara, have no page. That is not a judgement on their status: their UN votes do not exist, and the trade, conflict and migration sources treat them in different ways, so their scores could not be built with the same rules.
- The scores are built for UN member states. Palestine and the Holy See are non-member observer States: they cannot vote in the General Assembly, so they have no votes score and no index. The security score counts wars between states only, so the armed conflict over the territory of Palestine between the government of Israel and armed groups (UCDP conflict 234) is not in it; the Israel–Palestine page shows that conflict, and the trade in goods recorded by the Palestinian Central Bureau of Statistics (which the IMF data used for every pair does not have), as context. Neither enters a score.
- Trade agreements are those notified to the WTO (free trade agreements, customs unions, services and partial scope agreements), in force or signed. When a bloc such as the EU signs as one party, the agreement links its members with the other side, not with each other. Investment, tax and other treaties are not covered.
- Sanctions are the country regimes of the UN Security Council, the EU, the United States and the United Kingdom. Thematic lists (terrorism, cyber, human rights, chemical weapons) and other countries' own lists are not covered. A regime may target a government, officials or companies, so its presence says nothing about how far it reaches. Neither layer enters a score.
- Voting blocs and the voting map come from Nations on Record: Louvain communities on each country's eight closest voting partners per period, and a two-dimensional layout of each year's agreement. They are context, not scores.
- Trade covers goods, not services, investment or energy pipelines. Values are in current dollars.
- UCDP counts conflicts between states and organised groups with at least 25 battle deaths a year. Violence by groups without a government side, and conflicts between a government and armed groups based in another territory, do not count against that territory's government.
- Observers such as Palestine do not vote in the General Assembly, so some pairs have no index.
- Russia includes the Soviet Union and Germany includes West Germany, as in Nations on Record.
Sources and licences
- Nations on Record — General Debate speeches (UN General Debate Corpus, Harvard Dataverse), mentions, voting agreement (Voeten et al.; UN Digital Library), group memberships, trade (IMF International Trade in Goods). The terms of the original sources apply; UN voting data is for non-commercial use.
- UCDP Dyadic Dataset and Battle-Related Deaths Dataset, version 26.1, Uppsala University (CC BY 4.0).
- UN DESA International Migrant Stock 2024 (CC BY 3.0 IGO). UNHCR Refugee Population Statistics (CC BY 4.0).
- International football results 1872–present, Mart Jürisoo (CC0).
- Leaders: Wikidata (CC0), checked by hand for Nations on Record. Map shapes: Natural Earth (public domain), including its disputed and breakaway areas, drawn dashed (Nagorno-Karabakh and the 2014 lines in eastern Ukraine are left out as out of date). Prices: US consumer price index (CPI-U, Bureau of Labor Statistics, public domain), used only to show a pair's highest trade year adjusted for inflation; never in a score.
- Trade agreements: WTO Regional Trade Agreements database (rtais.wto.org), “Export all RTAs”, fetched 2026-10-25. Reused with attribution under the WTO’s terms.
- Sanctions: EU Sanctions Map (EU and UN Security Council regimes as applied in the EU; © European Union, reuse authorised), OFAC Sanctions Programs and Country Information (US Treasury; public domain), Current UK sanctions regimes (gov.uk; Open Government Licence v3.0); fetched 2026-10-25.
- Pictures: Wikimedia Commons, author and licence under each picture; resized and converted to WebP by us.
- Software and fonts on the pages: country flags from flag-icons (MIT licence), D3 (ISC licence), the fonts Inter and Instrument Serif (SIL Open Font License 1.1). All are served from this site.
- Voting positions, voting map and blocs: ideal points of Bailey, Strezhnev and Voeten and the bloc analysis of Nations on Record.
Loaded: 2026-10-12 20:33 UTC (Nations on Record) · scores built 2026-11-14 21:20 UTC.
Principles
We follow the principles of Nations on Record: one rule for every country; rate claims, not countries; context, not censorship. NationsRelations is independent and non-commercial, and not affiliated with the United Nations.