Summarise
Our local language model writes the record of each video of the week, checked against the transcript, then the syntheses of the day and of the week.
- 1Our local language model reads the whole transcript of the video.
- 2It writes the record, with its title, its summary and the claims made, in French and in English.
- 3Each record is checked against the transcript before publication.
- 4The week’s records give the synthesis of the day and of the week.
- Model
- Gemma 4 31B, on the project’s computers
- Guard
- names checked against the closed list; record held in case of doubt; synthesis sources checked
- Writes
- record per video, syntheses of the day and the week, section texts
A record for each recent video
Videos published since 26 September 2026 and transcribed within seven days of publication receive an automatic summary, written from the transcript by an open language model (Gemma 4 31B). It gives a descriptive title, four to six sentences on the content and the main claims made. Each claim is assigned to one of fifteen topics from a list specific to the summaries (security and justice, health, international affairs…), separate from the nine themes of the classifiers. The summaries of each day and of the week are then combined into syntheses by topic.
For daily and weekly syntheses, the model receives dossiers preserving the full summary, attributed claims and relevant transcript passages. The selection spans countries, orientations and channels within the model’s context limit. The narrative describes events and reactions, with video links directly in the sentences. A second automated reading checks the draft against the same dossiers; further checks validate cited identifiers and their association with channels.
The instructions require a neutral register and the attribution of every claim to its author. They forbid any addition from outside the transcript and any judgement. The list of people named only keeps those on the predefined list who appear in the transcript. A summary that seems to name a person who is not on the predefined list, such as a private individual, is withheld. Texts are not reviewed by the team; they describe what was said and do not constitute fact-checking of the reported events. Older videos, and videos transcribed more than seven days after publication, are not summarised.
The texts that open each section
On the home page, the trends page and the removed videos page, a short text opens some sections. It is written by the same language model, from the section’s data alone. It is renewed at most once a week, when these data have changed.
A text containing a figure that is not in these data is not published; the page then shows a sentence composed directly from the figures. Texts written by the model are labelled “Automatic text, not reviewed”.
The syntheses of the day and of the week, the records of removed videos (step 09) and the arguments by theme (step 08) follow the same rules, a neutral register, no addition from outside the sources, and no publication when a name could be a private person’s.