Laya vs Jev T-Rex
For each T-Rex game decision, a deterministic physics planner creates a text state and labels jump, duck, and run as safe or unsafe under the model’s answer timing.
Added to Jevfast
How it uses Jev
For each T-Rex game decision, a deterministic physics planner creates a text state and labels jump, duck, and run as safe or unsafe under the model’s answer timing. Jev chooses an action with a Choice question. The harness proposes the highest-probability action, then a safety shield can replace an unsafe choice.
What Jev decides
Illustrative source-derived request: choose the dinosaur actionExample answers · not a recorded Jev response · Source ↗
Question 1 · action
YOUR APP
INSTRUCTION
Choose the best safe action for the dinosaur.
STATE
Example state: “Dino runner game. 2 large cacti ahead, 96 px away.” The planner’s current safety result for each action is included in the per-option criteria.
JEV · CHOICE
- jump
- duck
- run
App workflow
- Observe the game
The game snapshot and timing estimate feed the deterministic planner.
- Label candidate actions
The planner simulates jump, duck, and run and supplies safe/unsafe descriptions for the present timing.
- Ask Jev and apply a shield
Jev returns probabilities; the highest-probability action is proposed, and the shield replaces it if unsafe.
- Repeat during play
While the round runs, the pilot asks from new game views and discards stale answers after key changes.
Why it is interesting
The project measures a hosted decision model in a live control loop while exposing the planner’s timing-aware candidate labels and safety interventions. Its scores describe the combined system, including the shield.