Pocket Pilot / Platformer

02 / A NEW PLAYGROUND

One leap at a time.

A Mario-style experiment. Walk, jump, and see every decision that gets you to the flag.

Original level · early experiment
No API key needed. Arrow keys / A D to move; Space / ↑ to jump.

Meadow run Level 01

Ready
Meet your next experiment.Reach the flag. Jump over crates, gaps and beetles.
Click Play, then use the keyboard or buttons below.
Hold to move · release jump before jumping again
Progress
0%
Coins
0 / 10
Decisions
0
Median wait
—
Game time
0.0 s

Observe → answer → advance. Simulation freezes during inference. Playback speed changes animation only; decisions always get the same frame budget.

Decision history

No decisions yet

Your exact inputs, answers and action outcomes will appear here.

WHAT WE ARE TESTING

Can a decision model learn when to jump?

The road asked for a lane. This game asks for one of six button combinations. First compare the original models with a visible scripted reference. Then collect trajectories and train a platformer-specific model.

Initial check: the scripted reference finishes the level. Clef’s first JSON run fell into the first gap; original Liquid initially chose to wait on JSON. These models are experimental here.

Your road fine-tunes are still in the Road lab. They predict left / middle / right; they have not been trained to jump. Liquid here is the original 600M model, not those fine-tuned weights.

What exists already?

typesafe-mario ↗ and mario-jev ↗ are community Jev controllers for NES Mario. They read emulator state, rather than screenshots. We have not confirmed a TypeSafe-authored Mario implementation.

SuperTux ↗ is a full open-source desktop platformer. This browser experiment has its own art, physics and single level; it is not the complete Nintendo game.

Hosted calls go directly to OpenRouter. Local Liquid downloads weights from Hugging Face and runs on your device. Each attempt stops after 200 decisions; exporting keeps the trace without your key.