Applied science: the CWM completes scientific cycles in public health and climate — against established baselines
The CWM's world scientist generalizes to epidemiology (5 capitals, 873 weeks) and wildfires (1998-2026): questions posed, forecasts risked, refutations accepted — and like-for-like comparison against established forecasting methods.
A world model only deserves the name when it faces the real world — with its incomplete data, long series, and questions without pretty answers. Version v0.10.6 takes the Camile World Model's world scientist, so far devoted to climate and finance, into two new real territories: epidemiological surveillance across five Brazilian capitals (dengue, 873 weeks of public data per city) and Brazil's fire hotspots between 1998 and 2026.
In each cycle, the model walked the complete scientific path: it posed the question, risked step-by-step forecasts — always without seeing the future — and accepted refutation when the data did not support the hypothesis. It also measured its forecasts against established time-series forecasting methods, under identical walk-forward conditions, with the numbers published.
The CWM's scientific cycle — question, prediction, falsification — now also runs on epidemiology and fire data, generalizing the capability beyond its original domains.
The model gained belief memory across datasets: a hypothesis refuted in one city reaches the next ones already carrying the learning — what fell does not resurrect as if nothing happened.
An open-arena comparison was published against established forecasting methods (classes of seasonal models widely used in industry), under the same test conditions for all.
Organizations dealing with public health, climate, or environmental risk series gain an instrument that does not merely forecast — it records how it formed and corrected its beliefs, and where its forecasts are (and are not) better than conventional methods.
Researchers and data teams can reproduce the complete cycles: every result comes with its verification records.
There is plenty of promise and little open science in the AI market. This version chooses the harder path: real questions, auditable public data, honest comparison with what already exists, and publication of results — including those that do not confirm expectations. That is how world intelligence becomes trustworthy infrastructure for decisions that affect people: with visible method, declared limits, and verification within anyone's reach.
In practice, a health agency can ask the model to read a municipal series and receive not only the projection, but the history of how often that kind of question has been tested, what was confirmed, and what was refuted. An environmental monitoring center can compare the model's performance against the method it already uses — on its own data — before adopting it.
The cycles measure forecasting and refutation on observational data — statistical association is not causation, and the model itself records when a question requires a laboratory the data cannot offer. Results are specific to the tested data and windows; they are not promises for any context.
Not every question has the expected answer — every question has an honest record. v0.10.6 is the CWM proving, on real ground and against established rivals, that it knows both things: how to forecast and how to recognize its own limits.