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.

1110-frame run · NTSC 60fps ✓ CERTIFIED REAL — byte-exact init, goal reached

Goal given to the agent: "reach the first ladder"

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. With no hand-written RAM map, the agent discovered on its own the bytes for Mario's horizontal position, his vertical position, and an on-ladder state flag — then walked Mario to the first ladder and climbed on. It consulted a vision/reasoning model only at decision points, and won at the first iteration.

Goal

Reach the first ladder — a real game objective

RAM map

Discovered, not hand-written (position + on-ladder state)

Won at

Iteration 1

Certification

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

The opening is Donkey Kong's own boot + intro animation (replayed from a reference movie so the trace certifies byte-exact); then the agent takes over, waits for the level to go live, and mounts the ladder. A static settle pause is trimmed from this clip for viewing — the linked movie is the full, unedited certified run.

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