Flappy Arena

A local web demo where Jev and GPT-6 Luna fly the same Flappy Bird course side by side. Every flap is a live model decision: each run starts on a fresh shared course, the birds take off together, and score, latest action, confidence and response time stay visible while they fly.

Clone the repo, run npm install, copy .env.example to .env and put TYPESAFE_API_KEY and OPENAI_API_KEY in it, then npm start. Open localhost:4173 and click Start the matchup; Node.js 22.19 or newer is required, and the keys stay on your local server.

A random seed gives both games the same pipe layout and physics, and the birds launch together after each model has made its first decision. Each model sees its own bird's projected position, velocity and next pipe, then chooses flap or wait; the projection accounts for expected response time.

Nothing blocks the game: the games keep moving at 30% speed while decisions are requested every 50 ms, several requests can be in flight, and a flap supersedes pending answers based on the old flight path. If an answer fails or arrives late, the bird keeps moving under gravity, and the browser shows what was applied, superseded or skipped.

The server writes Jev decisions to jev-logs.jsonl (latest 50) and GPT-6 Luna decisions to luna-logs.jsonl, while browser timing and game outcomes go to diagnostics.jsonl; those local files are git-ignored. The model logs are also readable at /api/jev/logs and /api/openai/logs.

No. The README publishes no aggregated win rates or accuracy figures; it is an arcade race you watch live, with per-run logs you can inspect. There is no CI benchmark or controlled trial behind the comparison, and outcomes depend on live API latency as much as on the models.

The README gives no cost figures; running it consumes both TypeSafe and OpenAI API usage from your own keys on the local server. npm run check runs the JavaScript syntax checks and game tests without playing a match.