Over the past year, coding agents have made game reverse engineering much faster. Matching decompilations, which rebuild a game's original binary byte for byte from C, used to take teams years. Now new projects, many of them built largely by agents, report matching most of a game in weeks or months.
A byte match proves the output is right, but not how it got there: recovered C on the original compiler, or C plus a patched compiler or per-file compiler settings. I was curious how those matches break down, so this personal side project rebuilds every matched function of public decomps (14 agent-assisted, 2 human-built for reference) on the released compiler with default settings, and counts how many come out the same.
- Ten of the 14 agent-assisted projects match at 99–100% on stock tools. Four fall below 99%, using patched compilers or per-file settings for part of their code; Pokémon Colosseum has the most, with 71.4% stock matches.
- Both human-built projects (Melee, The Wind Waker) match at 100%, counting community-shared compiler patches as stock.
- You can't tell these functions apart by reading the source; neither a static scanner nor an LLM classifier could. The rebuild is what finds them.
Write-up: WRITEUP.md · Run it yourself: RUNNING.md (Docker, no ROM needed) · Research log: findings.md