What was done
RovshanMuradov rebuilt the Revision 2017 winner Eidolon in Rust with identical frames and sound, then shrank a rebuilt version to 44,139 bytes. The first port took AI agents about 40 hours wall clock.
Eidolon by Poo-Brain won the PC 64k competition at Revision 2017. RovshanMuradov rebuilt it as a Rust program that reproduces all 27,000 frames and the whole soundtrack byte for byte under a controlled clock, then shrank a rebuilt version to 44,139 bytes, 27.6 percent smaller than the original with identical output. The project is framed as research into what frontier AI models do on a complete reverse engineering task.
How it was done
An agent team ran 15 to 17 October 2026: a coordinator that only planned, workers with fixed roles per step in separate git worktrees, and verification always by a different agent from the one producing the result, about 40 hours wall clock and 270 commits. The models used are listed with their token counts: GPT-6 Astra at high effort through OpenCode for the main work, GPT-6.1 Sol for a slice, and Claude Opus 5.5 as coordinator, over 746 million cached tokens in total. Analysis used Ghidra through the REA MCP server and objdump, unpacking by comparing memory dumps, and the original 64klang2 synthesizer was matched against its public source sample by sample.
Results
make verify checks every frame against captured SHA-256 values and make verify-audio checks the PCM data, and the size build reproduces the exact executable from a clean checkout. Documents explain how the original binary works and how the port is verified. Native Windows playback and the 32 KiB target stay open. The demo shaders, music and animation data remain the work of their listed authors, credited in the repository; code is MIT or Apache-2.0.