Donkey Kong, solved by an AI agent

No hand-written RAM map. Given only the ROM and a goal in plain English, the agent discovered how the game works and beat it.

720 frames · 12s · NTSC 60fps ✓ CERTIFIED REAL — byte-exact init, goal reached

Goal given to the agent: "walk right until Mario's horizontal position reaches 160"

A pure-Python NES simulator (its 6502 CPU passes all 8,991 nestest instructions) drives a go-explore solver loop: perceive → model → plan → act → observe → learn. The agent discovered the RAM bytes for Mario's position on its own (delta-shape analysis under controller perturbation), consulted a vision/reasoning model only at decision points, branched on savestates, and won at iteration 21.

Approach

Go-explore + LLM at macro boundaries only

RAM map

Discovered, not hand-written

Won at

Iteration 21

Certification

Byte-exact to FCEUX reference within the trusted horizon; gameplay sim-trusted

The first ~9 seconds are Donkey Kong's own boot + intro animation (replayed from a reference movie so the trace certifies byte-exact); the agent's controlled gameplay begins once Mario becomes controllable.

Download the raw input movie: dk_solve.fm2 (FCEUX format — replays this exact run).