Glaurung Recompiles Its Own Reverse Engineering

Field notes

The project stores analyst edits and inferred information in SQLite, recording provenance so manual names outrank automatic ones. Its validation corpus spans sorting, hashing, SHA-256, and Duff’s device, with each recovered function compiled into a shared object and exercised beside the original. The visible baseline reports 83.9% behavioral agreement across 3,449 lanes, while separately exposing weak spots such as Rust optimization lanes and reads from unwritten locals.