Community · E2 · artifact verified
Classify dino obstacles, let physics time them
An autonomous bot that plays the Chrome T-Rex Runner to a thousand points by asking Jev which action each obstacle requires - jump, duck, or run - and letting a measured physics model decide exactly when to press.
01 · Role in the system
What Jev does here
The page runs the real open-source Chrome dino game with its internals exposed, and for every obstacle sends its type, height, width, speed, and gap to the System One endpoint as a choice question with three options. The README is explicit about the division of labor: the decision comes from Jev, but the timing comes from a physics model derived by measuring the dino's actual jump arc - a fixed 34-tick flight - and solving for the trigger distance that centers an obstacle's danger window inside that arc, because a semantic-judgment API with a half-second round trip is not built for sub-second reflexes. A local proxy fronts the official endpoint for the browser.
02 · Control boundary
Where Jev sits
Per-obstacle three-way action choice; a measured jump-arc physics model owns trigger timing; the proxy keeps the official endpoint browser-reachable.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- Tuned to the canonical Chrome dino physics; reskins or modified gravity would need re-measuring the arc.
- The thousand-point target is the README's demonstrated ceiling, not a benchmark of limit performance.
- Each obstacle costs a network round trip, so very dense clusters rely on the physics fallback between calls.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- kevinwindisch ↗Community · github · public · checked 2026-09-25