BUILD / PLAY / VERIFY
A game built by Arbiter.
A player called Sentinel.
From generated physics to measured decisions.
Inspect the build, watch actual trials, or take the controls.
Loading verified engine and recorded trials…
The experiment has a visible result.
Sentinel passed zero pipes on all three seeds, both with and without exact simulator forecasts. It is not yet a reliable game controller. We retain every run, including failures, rather than substitute scripted actions.
| Controller | Seed 7 | Seed 17 | Seed 29 |
|---|---|---|---|
| Sentinel | 0 | 0 | 0 |
| Sentinel + forecasts | 0 | 0 | 0 |
| Scripted baseline | 5 | 1 | 5 |
| No action | 0 | 0 | 0 |
Pipes passed, not win rates. Seed 29 baseline reached the decision budget; other runs ended in collision. Forecasts come from exact simulator physics, not a learned world model.
Code generation.
Verified by execution.
Arbiter generated this game's physics engine. Independent tests caught a boundary-collision bug. Arbiter repaired it, and all six checks passed.
The specification, tests, controller integration and interface were supplied by the development harness. This demonstrates a bounded generate–test–repair sequence, not unrestricted autonomous software development.
Explore decision tasks ↗- Actual served model
- arbiter-flash-27b
- Generation calls
- 2 — initial build + repair
- Verified behaviors
- Determinism, physics, collisions, score-once, input validation, boundary contact
- Generation time
- Loading…
Times are observed local generation times, not a service guarantee. This is a Flappy-style prototype; Doom is not demonstrated.