How Mecapress works

A story groups the events of a same case : the original fact, then the announcements, reactions, decisions and what follows from them. Each sub-event keeps its own category, severity and sources; they are never merged.

After each collection, new events are compared with the events of the last 7 days and with open stories: places, people and organisations in common, rare words in the title, same brands or same countries. Plausible matches are then submitted to an AI, which only groups them when they share the same triggering fact — not just the same topic or the same city. A story only exists from 2 events.

The story title and summary are written by AI from its sub-events, and updated when a sub-event arrives or changes (at most once an hour). The story severity is the highest of its sub-events. After 7 days without activity, the story is closed: it can still be read but takes no new sub-events. An administrator can detach or move an event or rename a story; the automation then respects these choices.

The timeline “Hottest” rates subjects according to activity in the chosen period (6 h, 24 h, 3 days or 7 days). A story counts only once there, with the activity of all its sub-events.

• Activity : each article published during the period counts, with more weight given to recent ones (its weight is divided by e ≈ 2.7 every third of the period);
• Diversity : activity is multiplied by 1 + 0.6 × ln(1 + number of distinct sources), so that a subject covered everywhere ranks ahead of one hammered by a single outlet;
• Severity : multiplier from 0.75 (severity 1) to 1.75 (severity 5).

The displayed score is scaled to 100 for the top subject. The trend compares the number of articles to that of the previous period of the same duration: up by more than +25%, down by less than −20%, “new” if there was nothing before.

For active stories covered by at least two outlets, an AI lists the key facts reported (who, what, where, tolls, amounts, decisions) using only the headlines and extracts of the collected articles, and records which articles state each one. It consults no other source and adds no facts.

The classification is not left to the AI: a fact is “confirmed by several sources” when at least two independent outlets report it. Outlets with the same owner (for example titles of the same regional press group) count as a single source, since they often share the same wire copy. Otherwise the fact is listed as “single source so far”.

“Sources disagree” flags incompatible versions of the same fact at the same time (a toll, an identity, a place) coming from different outlets. A change over time (a rising toll, a denial) is not a disagreement: it appears in the dated updates.

The cross-check is redone at most every 3 hours when new articles arrive, hottest stories first, using the most recent articles (at most 3 per outlet; posts on X are not counted). “Confirmed” means “reported by several newsrooms”, not “verified on the ground”: several outlets may rely on the same original source.

For the hottest stories, the 2 or 3 central claims of “What is established” (the two most widely confirmed facts and the main single-source fact) are checked on the web by an automatic search service, which looks for official sources first (prefecture, ministry, prosecutor, police, local authority, institution), then for established media.

Each claim gets one of three neutral verdicts, with the sources cited: “confirmed” (established by an official source or several established media), “not confirmed at this stage” (nothing solid establishes or contradicts it), or “contradicted by an official source”. The last verdict is only kept when an official site is actually cited; otherwise the claim stays “not confirmed”. A story with a contradicted claim shows a “Denied” badge. A verdict requires a probability of at least 60%. Claims are re-checked after 8 hours (24 hours when confirmed), within a daily budget.

Newsroom fact-checks. Mecapress also reads the fact-checking sections of AFP Factuel, Les Décodeurs (Le Monde), CheckNews (Libération) and franceinfo “Vrai ou faux”. Each article is compared with the stories through semantic memory, then an AI checks that it verifies a precise claim about that very case, not just the same topic. The story then shows “This claim was checked by…” with a link to the article and the newsroom’s conclusion.

These checks do not replace the cross-check of the press: they add an outside look. An automatic search can miss a source or be overtaken by events; the date of each check is shown.

The hottest stories get a “Background” box, set apart from the news. Its aim is to help readers understand rather than just follow events, in a neutral tone: no sensational wording, no causal conclusion, no political judgement and no prediction. Each part is only shown when reliable data exists.

Reference points (comparable past events, from Mecapress’s archive and Wikipedia), key terms (definitions from Wikipedia, with the matching Service-public.fr page) and what happens next (usual steps, without prejudging the outcome) are written by an AI only from the numbered sources provided. An automatic check then drops any item that cites no source, contains a number missing from its sources, or uses sensational wording. Useful resources (help numbers) come from a fixed list of official services.

The box is refreshed when the story changes (at most every 12 hours), within a daily budget. The administrator can hide or regenerate it.