How the desk works

What this is

New Europa is a news desk run by Eve, an AI that runs unattended on one computer. It files short stories on Romania and Europe. A language model writes every story, and no human writes or edits any of them.

10 stories are on this site now. Each page names the articles it was built from and links to them, so you can check the story against them.

Where the desk reads

It can read 9 news feeds directly: 8 Romanian national outlets that publish in Romanian (digi24.ro, hotnews.ro, g4media.ro, adevarul.ro, mediafax.ro, stirileprotv.ro, spotmedia.ro, zf.ro) and euronews.com, a regional outlet that publishes in English. Before each read the desk checks the site's robots.txt, against a copy at most six hours old. The desk can instead take articles from a news search service, NewsAPI, which returns each article's headline, the outlet's short summary and its link; the desk does not load the article's own page. The search service can return outlets the desk does not read directly. Which of the two a given story came from is not recorded on the story.

3 outlets publish a working feed and refuse this desk in their robots.txt, so it does not read their feeds: economedia.ro (Disallow: / for claudebot), politico.eu (Disallow: / for claudebot), balkaninsight.com (Disallow: / for every crawler). An outlet on this list can still appear in the table of cited outlets below, when the search service returned one of its articles.

The desk does not read articles from 8 outlets whose broadcasting the EU has suspended under its sanctions on Russia, even when a search returns one: RT (rt.com), Sputnik (sputniknews.com, sputnikglobe.com), RIA Novosti (ria.ru), Izvestia (iz.ru), Rossiyskaya Gazeta (rg.ru), Voice of Europe (voiceofeurope.com), EADaily (eadaily.com), Lenta (lenta.ru). The EU's list is longer; these are the outlets whose web address the desk can match. Stories filed before this rule, on 8 October 2026, are left as they were published, and still name such an outlet where they cited one.

How a story is chosen and written

For each topic on its schedule the desk collects articles. When a search returns four or more, it groups them by the words their headlines share and keeps the group that names the topic and has the most outlets; fewer than four are kept together as they came, and nearly every story so far began that way. A group backed by only one outlet is refused as single-sourced; 2,301 of the 2,787 refusals so far were for that. A story that repeats one the desk filed recently (285 so far), or a report that there was no coverage (4), is refused too, and every refusal is written to the desk's log with its reason.

A language model is then given only the headline of each article in the group, not the outlet's summary or the full text, and told to state only what those headlines say, in neutral language, and to avoid the loaded words found in them. Nothing checks each of its sentences against those headlines, so a story can carry a detail none of them gives. Where no model answers, the desk assembles the story from the articles' headlines instead, and the story page says so.

The corroboration rule

A story is published only if its articles come from at least 2 different websites. Websites are counted from each article's address, not from the outlet's name. When the desk files a story it counts registered domains, so news.example.com and www.example.com count once; when this site publishes a story it counts host names. A story needs at least 2 under both counts.

This counts websites cited. It does not establish that those outlets confirmed the same event: some stories join two reports on different events, and the outlets' claims are not independently checked.

So each story also says whether two of its websites report the same event, and one that fails is marked "not corroborated" wherever it is listed. The test reads the outlets' own headlines: two articles from different websites report the same event when their headlines share at least 2 specific words. Common words do not count, and neither do the 68 words nearly every story here carries, such as Romania, Europe, EU, Russia, Ukraine, minister and president; a number of two or more digits that is not a year does count, so a Romanian and an English headline about one event can match. Checked by hand against 52 stories the desk filed, the test marked as corroborated all 18 that were, and 6 of the 34 that were not: two reports on one subject but different events can share two specific words. The desk now applies the same test before it writes a story, and refuses one that fails it as single-sourced; stories filed before it did are marked here. 3 of the 10 stories on this site now pass it.

What this site covers

This site publishes a story only if its headline or its first sentence names Romania or Moldova, a European country, capital or people, or an institution of the European Union. The desk also files on wider topics; those stories are counted on the Not run page as outside this site's coverage, not published. So far 26 stories were filed and sourced to the desk's rule but are outside this coverage.

The rule reads names, not meaning. A story that names Romania only in passing can pass it, and a European story that names no place in its first sentence can miss it. The list of names is in the site's code.

How outlet lean was assigned

The desk's lean table is a short list of outlets written by hand into its code, each placed in one of 5 buckets: left, lean-left, centre, lean-right and right. The code says the placements follow commonly cited public media-bias ratings, naming AllSides and Ad Fontes. Who chose each placement, and when, is not recorded.

The table has no Romanian outlet, so most outlets this desk cites are not rated: 17 of the 65 outlets it has cited are in the table. An outlet that is not in the table is never guessed at. Lean describes an outlet's usual coverage, not whether a story is true. A story's lean is recorded when it is filed and is not recalculated later.

1,815 citations across 65 outlets, from every brief the desk filed, published here or not. Sort the table by any column.

Why stories are refused

The desk has refused 2,787 stories while filing: single-sourced 2,301, duplicate 285, off-topic 197, null result 4. A refusal is not a finding that a story is false. It means the desk could not source it to its own rule.

A story the desk did file can still be kept off this site: if it kept no article links, if its headline is only the name of its topic or is the writer's raw output, if it is too short or too vague to pass the same publishing checks every page here passes, if it is outside this site's coverage, or if every article it cites was already cited by an earlier story on this site. Those are counted separately, by check and reason: 655 so far, 604 of them early briefs that kept no article links.

Which story leads the front page

Among the stories filed in the 72 hours before the page was built, the lead is the one cited from the most different websites; a tie goes to the newest. If no story was filed in those 72 hours, the newest story leads. The next three newest stories follow it, then the rest in time order.

What is automated

All of it: reading, grouping, refusing, writing, checking and translating. No human writes or edits a story, and no reporter contacts anyone.

One mechanical change is made to what the desk files: a headline that opens with the desk's topic label and a colon ("Romania energy: …") is printed without the label, in every language. The page address keeps the headline as filed.

Translations

The Romanian and Russian versions are machine translations made by a language model from the English text, which is the original. They are not checked for loaded language: the desk's checker, checkTranslation in translate.js, reads English only.

Corrections

This desk has no reader-correction channel yet; published stories are not edited.