AI & product
AI with an inspectable evidence boundary
The different jobs of extraction, multilingual summaries, and WOUT’s source-backed web assistant.
The quality of an answer depends on the records available before the question is asked. A fluent answer cannot repair missing evidence.
Separate the model’s jobs
Extraction helps read an incoming item. Summarization helps people understand it. Translation helps make it accessible in Kreyòl, French, and English. Each operation can introduce errors and needs an appropriate evidence boundary.
WOUT labels machine translations. Model extraction confidence must not be presented as a percentage certainty that an event happened.
Ask WOUT uses a bounded news window
The web assistant receives WOUT’s published news from the preceding 48 hours. Its instructions require item-level citations, attribution of claims to the source, recognition of retractions, and a plain response when the available feed cannot answer a question.
This is an instruction-based grounding boundary, not a guarantee that every generated answer is correct. Readers should open the sources. The assistant is not a complete view of current road conditions or an emergency service.
This site’s automatic analysis is calculated
The analysis page calculates event counts, evidence-state distributions, category distributions, represented areas, and recent source-linked news from public records. It does not ask a model to invent a narrative or predict what will happen next.
Clear calculations make the report reproducible: the sample limit, time window, exclusions, and source failures are displayed with the result. Future model-generated commentary would need its own evaluation and publication policy.
Based on the WOUT implementation reviewed October 11, 2026. Rules and product availability can change. Examples are explanatory; current records appear in the public analysis.
Inspect the latest public calculation