Kristian Fagerlie explores Jev as a decision component rather than a chatbot that writes a free-form answer. A caller supplies structured observations and a limited set of allowed actions; the model returns probabilities over those choices. A game demonstration uses health, ammunition and positioning to choose among moving, taking cover, flanking and reloading, while the game engine still executes the action.
The experiments then ask whether the same interface can interpret ambiguous news and simulated market conditions. Fagerlie contrasts a hypothetical price-moving football update with a published lineup that already contains expected information. He also tries semantic questions and a toy order-book simulation, treating the model as a way to encode fuzzy conditions without writing an expanding collection of rigid rules.
Fagerlie repeatedly separates confidence from correctness and acknowledges that the market experiments have not demonstrated a useful edge. The live-data integration is exploratory, and shadow or paper testing is proposed as further work rather than reported as a validated profitable strategy. The useful engineering lesson is the constrained decision interface and the need to measure outcomes, not to mistake a confident output for an accurate prediction.
Watch on YouTube
