Jev obstacle action choice reads the obstacle and its clearance to select one maneuver. The six samples test ground obstacles, different bird heights and unresolved geometry.
A dinosaur-game controller can separate the question of which maneuver fits an obstacle from the question of when to press a key. This original demonstration handles only the first question. The context describes the obstacle and the relevant clearance in words. Jev returns Jump, Duck, Run or Need more context. The downloadable consumer records a pending maneuver without controlling a character.
The API-collected source describes a game experiment that also separates narrow model choice from code-owned physics and timing. We use that division to build six synthetic descriptions, not to reconstruct the game or verify its recorded performance. A cactus, a bird at head height and a clear lane need different evidence. A bird with unspecified height has no justified maneuver under the supplied rule.
These fixtures test a decision contract. They do not test frame rate, reaction time, survival duration or how a moving obstacle is detected. The actual API round trips appear with their saved responses. A game would still need to schedule actions against its own physics. Unknown geometry and contradictory notes remain explicit outcomes instead of producing an invented clearance estimate.