How we keep this honest

This is the rule the whole record is built around, and it is enforced in the data model rather than left to an author's judgement. A casualty figure in an active war is not a fact you look up; it is several institutions counting different things and arriving at different answers. So a figure here is never a bare integer. It carries a range where the truth is a range, the counting BASIS (what is actually being counted, which is where most public disagreement really lives), a confidence, and every competing published figure side by side with its source and date. Where credible sources disagree, we widen the range to span them and label it disputed. We do not average them, and we do not pick a winner. Gaza is the clearest case: UCDP records incidents across Israel, Gaza and the West Bank, while the Gaza Ministry of Health count covers Gaza alone and does not separate combatants from civilians. Those two numbers are not reconcilable into one, so we publish both and say why they differ.

The most dangerous number in a conflict record is a confident zero. When the United States struck Fordow, Natanz and Isfahan in June 2025, Iran said the sites had been evacuated and no source published a casualty figure for those specific strikes. The honest record of that is UNCONFIRMED, not zero: a hard zero would conflate two different facts, the deaths those strikes caused (unknown) and the US military personnel killed (none). The same discipline applies everywhere. Where nobody credible has published a figure, the interface says unknown and renders it differently from a measured value.

Whether a situation is an international armed conflict between states, a non-international one between a state and an organised armed group, or a military occupation is a legal question with real consequences, and the political framing of a war and its legal framing are usually different. Rather than make that call on our own authority, we defer to the Geneva Academy of International Humanitarian Law and Human Rights, whose War Watch project classifies armed conflicts as a profession and publishes its reasoning. Each conflict page lists the classifications that belong to that record, named exactly rather than matched by keyword, with their current titles pulled live from War Watch and linked back to their analysis. This routinely reveals what a single headline label hides: most of these theatres contain several legally distinct conflicts at once, with different parties and different applicable law.

A conflict in this record has parties. There is no field for who started it, no field for who is right, and none will be added. Each party's entry states what that party does, factually. Where a state fights alongside a co-belligerent, we name the co-belligerent rather than folding it into 'government forces', following the Geneva Academy's own practice: Mali supported by Russia's Africa Corps (the Wagner Group until 2025), Burkina Faso supported in places by VDP militia units. Naming who is actually fighting is more informative, and less editorial, than a tidy label.

Several conflicts here have no name that every party accepts, and silently choosing one party's term is taking a side in the dispute the record exists to document. Russia calls its invasion of Ukraine a special military operation. Israel, the United States and Iran each named their June 2025 operations differently. Almost every term in Israel and the Palestinian territories is contested, including the names of the parties and the status of the territory. We use a neutral descriptive name, record each party's own term where it is relevant, and flag the contest on the page rather than burying it.

The map is built from OpenStreetMap data (served as vector tiles by OpenFreeMap), and it renders disputed boundaries as a distinct dashed outline rather than resolving contested borders into settled ones. Drawing a contested border as final would take a side in exactly the territorial questions several of these wars are being fought over.

Conflict-level figures come from the Uppsala Conflict Data Program's Georeferenced Event Dataset: individually geocoded, dated and sourced incidents of organised violence, each carrying UCDP's own low, best and high fatality bound. We chose it for three reasons. It is produced by an academic programme that is party to none of these conflicts, which is the strongest neutrality position available. It is licensed CC BY 4.0 with redistribution explicitly permitted, so our own record can stay open. And its uncertainty is native: because every incident already ships a range, we inherit the honest uncertainty rather than manufacturing it. The incident record currently runs to 30 Sep 2026, combining UCDP's annual release with its provisional candidate releases (a reviewed release each quarter, monthly releases in between), and the provisional share of each total is disclosed. Where UCDP can place deaths only within a period or an area (weeks of fighting, a province, a whole country), the timeline spreads them evenly across that period and the incident says so, rather than stacking them on one day or implying an exact spot.

UCDP counts DIRECT deaths from organised violence in incidents it can source from public reporting. It does not count indirect deaths from disease, starvation and collapsed health care, which in several of these wars are believed to exceed the direct toll, and it cannot count violence that never produced a sourceable report. Every figure here is therefore a documented floor, not a total, and the interface says so with the words 'at least' rather than presenting a floor as a sum. In Sudan, a peer-reviewed mortality study of a single state found more violent deaths than event-based datasets recorded for the entire country over the same period. That gap is the point: an event count must never be labelled 'the death toll'.

A cumulative toll mostly measures a war's length: a war with a small total can be recording more deaths this quarter than one whose total is ten times larger. So beside each total, the record draws the deaths UCDP recorded per month over the last two years it has coded, and sums the latest three whole months (Jul to Sep 2026), comparing them on its own page with the three before. Active wars are listed in that order, deadliest now first. These months are always UCDP's, for every war, even where a war's headline toll comes from another counter (the UN's for Haiti, a human rights group's for the Iran-US war), because a ranking that mixed counters would rank the counters rather than the wars; where they sit below a headline, the page says whose count each is. They carry UCDP's own limits: a documented floor, coded weeks behind the war, and blind to violence its definitions do not cover.

Where a record shows how many people have been forced from their homes, the figure comes from UNHCR's official population statistics for the end of 2025, licensed CC BY 4.0: the refugees and asylum-seekers from the country plus the people displaced inside it. It counts people at a moment, not movements, so anyone who has gone home is no longer in it. UNHCR records where people are, not why they fled, so a country's figure covers every conflict and persecution there, and each record says so beside the number. Where a war has a more specific and more recent count, such as IOM's for Lebanon, that count is used instead and named.

ACLED is an excellent conflict dataset and we do not use it. Its terms require a paid licence for commercial entities, forbid reproducing, republishing or creating derivative works, and state that supplementing or reorganising the content or surfacing it through your own dashboard is not sufficient to escape that. A record that publishes itself as open data cannot honour those terms, so we do not ingest it, and we do not ingest it indirectly through aggregators that repackage it either. We also decline sources that are non-commercial-only or that publish no licence at all, however convenient. Where such a source is worth reading, we cite and link it as journalism rather than absorbing its data.

For the one conflict with a full event-level dossier, the June 2025 Iran-Israel-US war, nothing appears on the map without a citation. Each event carries the name and URL of the report it came from, tagged confirmed, reported or disputed, and every marker sits at the real coordinates of the place it describes rather than being approximated to the nearest capital. Confirmed means multiple credible sources or a primary record agree; reported means a single credible source; disputed means the core facts are actively contested and the note explains who claims what.

Each record states exactly what period it covers, and we guard hard against contamination between conflicts that sources routinely conflate. A war that began after January 2020 is counted from its own first day, so the Sudan civil war's toll starts on 15 April 2023, not with the violence in Sudan before it; a war that began earlier (Yemen's in 2014, Syria's in 2011) is counted from 1 January 2020, when this record's window opens, and its figure says so beside the number. The clearest example of contamination: the June 2025 Iran-Israel-US war documented in our deep dossier is a DIFFERENT conflict from the later Iran-United States hostilities, and search results mix the two constantly. The Geneva Academy keeps them as separate classifications, and so do we. Incidents are assigned to the war UCDP codes them to before the country they happened in, so a war fought across several states has one record and none of its deaths is counted inside another country's war. A record that quietly absorbs the wrong war's figures is worse than no record.

Each record opens with a short statement of what the war is: who is fighting whom, since when, at what scale, and where it stands. That holds for months. The week's news does not belong in it, because news ages in days and needs its date and its source beside it. So every recent development is listed on its own, dated to the day it happened rather than the day it was reported, newest first, and linked to the report it rests on, with contested claims attributed to whoever makes them. A reader can see at a glance what is new, and check each line against its source. A development is also drawn on the globe, on its war's timeline, where its own report says it happened: a strike, an attack on a ship, a battle, a capture or a displacement is placed at the town or site the report names (a diamond), or, when the report names only a region, a province or a sea, as an area (a dashed ring), never as a pinpoint. A statement, a talk, a vote or a report is not placed, because where something was said is not where the war happened. The coordinates come from a gazetteer (OpenStreetMap, Marine Regions, or this record's own chokepoints), and an automatic check fails any place its development's text does not name. Because the incident data is coded weeks behind the war, a war's timeline runs past it to the day the record was last verified; that stretch is hatched as not yet coded, the death count stops at the last day the data covers, and what the globe shows there is the dated developments.

The historical dossier (72 events) and the conflict registry (16 conflicts, 762 distinct sources) are version-controlled in code, so they are auditable and cannot silently change. The whole verified record is downloadable as JSON and CSV under CC BY 4.0. If you think a figure is wrong, you can check our arithmetic against the sources we name.

This is a living record and we will get things wrong. If you can show, with a source, that an event, a figure, an attribution, a classification or a coordinate is inaccurate, we will correct it and note the change. Accuracy is the entire value of this project; a correction is a feature, not an embarrassment.