About this article
A guide to one court judgment, built from the judgment and the dissent of six of its judges, and nothing else.
What it is
On 19 November 2026 the International Court of Justice gave its judgment in The Gambia v Myanmar. It runs to 191 pages and 646 numbered paragraphs. This article sets out what the court decided, where and when the events took place, and what the court ordered, for readers who have not followed the case.
The top of the article gives the ruling in six points. Below that, each section starts with one sentence that carries its main point. Panels, the map of nine places and the table of votes are there for readers who want more.
The one rule
There are two sources, both published by the court on the same day: the judgment, and the joint dissenting opinion of the six judges who voted against the finding of genocide. Nothing in the article comes from news reports, from earlier knowledge of the case or from anywhere else. Every factual statement ends with the paragraph number it rests on, and says which of the two documents it means.
The two are kept apart. What the court found is given as the court’s. What the minority said has its own section, and one point in the summary at the top.
That rule has a cost. Sixteen other opinions and declarations by individual judges are not covered. The article does not say where The Gambia is, or what has happened in Myanmar that the two documents do not mention.
How it was checked
Three ways, in this order.
- By machine. A script reads every block of text in the article and looks up the paragraphs it cites. Each figure in the block must appear in those paragraphs. Each quotation must appear there word for word. A block with a figure and no citation fails. The table of votes is built from the judgment’s final paragraph and compared with it again on every run. The first run found 38 problems. It now finds none.
- By two independent readers. A script cannot tell whether a sentence is a fair summary. So two AI reviewers, working separately and without sight of how the article was written, read it against the judgment sentence by sentence. They were asked to look for anything overstated, anything attributed to the court that was only a party’s argument, and anything favourable to Myanmar that had been left out. They found no factual errors. They found 14 places where a qualification had been dropped and about 30 smaller points of wording. All were corrected.
- By the limits of the method. A few numbers are counts made for this article, such as “four of the five acts”. They are marked in the page’s code and listed by the script each time it runs, so that none passes as the court’s own figure.
No check of this kind replaces a lawyer or an editor reading it. It has not yet had that reading.
The maps and drawings
Coastlines and borders are from Natural Earth, which is public-domain map data. The judgment prints five sketch-maps. The positions of villages, and the lines between the three townships, were read by eye from the first of them, using the latitude and longitude lines printed on it. They are approximate.
In the drawing of people, each dot stands for ten. The dots are scattered at random inside each township, weighted by the Rohingya population the court gives for it. They do not show where anyone lived. Where the court says “more than 700,000” or “at least 10,000”, the drawing uses 700,000 and 10,000. The dots that move end up spread over the nearest part of Bangladesh. The judgment says the refugee camps are at Cox’s Bazar but gives no position for them.
In the drawing of villages, each block is one village. The counts are the court’s. Which blocks fall is arbitrary.
Testimony
The judgment quotes survivors describing the killing of children and rape. The main text says plainly what the court found. The quoted testimony sits in panels that are closed until the reader opens them, each labelled with what it contains.
How it is made
Plain HTML and CSS, with one JavaScript file. The map, the dots and the villages are drawn with three.js. All of the words and numbers are in the page itself, so the article reads in full without JavaScript, by keyboard and with a screen reader. Animation is switched off for readers whose devices ask for reduced motion. No fonts, trackers or other files are loaded from outside.
The article was written and built with Claude, Anthropic’s AI model, working from the judgment under an editor’s direction.
Diary
- Made public. The article and its code were published on the evening of the judgment. It went out without a lawyer’s read, which is why the page says so above and asks search engines not to index it. The two court documents are not copied here. They are on the court’s own site.
- The dissent added. The first version could not say why six judges voted against the finding of genocide, because their opinion was a separate document. That was a real gap in an article about a nine-to-six ruling. The joint dissenting opinion is now a second source, with a section of its own and a point in the summary. The checking script was extended to read it, so a citation marked “dissent” is checked against the dissent’s own paragraphs. A third independent reading was run on the new section. It found no factual errors and 13 points to correct. Two mattered most: the first draft had left out the minority’s argument about mental harm and its argument about what “destroy” means, which made its case look weaker than it is. Both are now in. The court’s website blocks automated downloads, so the document was saved by hand.
- First version. Built on the day of the judgment. The whole judgment was read before any copy was written. Four choices were settled first: a sober, third-person voice; a look of its own, with no newspaper branding; testimony behind panels the reader opens; and real coastlines, with village positions taken from the court’s own sketch-map. The article has five scenes that respond to scrolling (the region, the three townships, the days of August 2017, the people as dots, the villages as blocks), a map of the nine places the court examined, the legal test laid out act by act, and every judge’s vote on all 13 questions.
- Checked and corrected. The checking script and the two independent readings described above were run before the first version was shown to anyone. The corrections mostly restored the court’s own caution: that it could not fix the exact number of dead at Min Gyi, that the birth limits were not confined to Muslims on paper, that the commander-in-chief’s speeches were judged “as reported”, and that incitement of the state’s own forces is absorbed by the finding of genocide and was not dismissed.