Most media datasets are frozen the day they are published. This one keeps watching. New climate coverage is picked up as it appears, then extracted, segmented and annotated by the same validated chain you just followed. From there a local language model running on our own servers writes the short bilingual summaries you see across the observatory, and our detection method surfaces the events and cascades as coverage converges. The full cycle runs every day, so the corpus — and everything it feeds — stays current without anyone touching it.
Every new article follows the same steps, automatically, every day — from raw coverage into the corpus, then to bilingual summaries written by a local language model running on our own servers and the events and cascades it surfaces.
Once articles land in the corpus, a local language model runs on our own servers to write the concise CCF summaries — per article, per event, per cascade and per region — in English and French. Full article text is never redistributed; the model reads the source and returns a brief.
Annotated articles are compared by semantic proximity. When several outlets converge on the same occurrence, our method consolidates them into an event, then titles, summarizes and locates it. See how events are detected.
Our open scientific framework then traces how a frame travels from one outlet to another. Each retained cascade is enriched in both languages before publication. See how cascades are detected.
