GitHub - neuromancer/rats-re

GitHub

5 min read Original article ↗

Rats! source reconstruction with local LLMs

This repository is a work-in-progress reconstruction of the source code for RATS.EXE, the original Windows version of Rats! (1994) by Sean O'Connor. It builds a Win32 executable with Microsoft Visual C++ 4.1 under wibo and can be tested in DREAMM.

The experiment explores how far local LLMs can take source reconstruction on a small but realistic codebase rather than a toy example. Match fidelity remains uneven, which is useful here: the aim is to present a realistic view of the process and its current limitations, not to hide difficult functions.

Function reconstruction is restricted to local LLMs. Codex and Claude are used for repository setup and benchmark orchestration, but they are not allowed to write or edit the reconstructed source.

Unless noted otherwise, all experiments used Qwen3.8 27B BF16 on a 2024 MacBook Pro with an Apple M4 Max (16-core CPU, 40-core GPU) and 128 GB of memory. The approach works—imperfectly, but usefully: the 36 retained functions took an average logged time of 2m 32.1s each to reconstruct.

Setup

Clone with submodules, or initialise them after cloning:

git submodule update --init --recursive

The original executable is not included. Download the free Windows version from the official Rats! page and place its RATS.EXE in the repository root to use the original-executable and binary-comparison targets.

Common commands

make                 # build out/RATS_RE.EXE, .map, .obj, and .asm
make toolchain       # download, extract, and verify Microsoft Visual C++ 4.1
make run             # launch the rebuilt executable in DREAMM
make run-original    # launch the preserved original RATS.EXE
make test            # bounded DREAMM smoke test of the rebuilt executable
make test-original   # bounded DREAMM smoke test of the original executable
make debug           # start the rebuilt executable in DREAMM's debugger
make report          # report similarity for every reconstructed function
make compare-func FUNC=SaveHighScores ADDR=00409092

The checksum-pinned MSVC 4.1 archive, the required MSVC 4.1 libraries, wibo's compatible msvcrt40.dll, and DREAMM 4.0x21 are downloaded on demand. They remain ignored local build dependencies and are not committed.

Reconstruction workflow

The checked-in ghidra/ directory contains assembly and decompiler exports for all 177 internal functions. Assembly is the comparison authority; decompiled C is only a semantic seed.

Install binary-recons once, then run it from the repository root:

python3 -m pip install -e /path/to/binary-recons
binary-recons --next-function           # reconstruct the next safe missing target
binary-recons --address 0x409092        # reconstruct or improve a specific target

The tool discovers Qwen automatically in the standard Hugging Face cache; BINARY_RECONS_MODEL_PATH can override it. binary-recons.toml keeps automatic selection inside the game-code ranges, starts and stops llama.cpp, and defines the transactional files and binary-comp command. Each bounded run seeds one function from Ghidra, asks Qwen for focused compile or assembly-diff repairs, and retains the best safe compiling candidate. Use --target-score 95 for a deeper pass or --dry-run-prompt to inspect the prompt without loading Qwen.

Measured local-model runs are recorded in docs/MODEL_RESULTS.md.

Model results: 36 retained functions and 3 deferred addresses

Most retained candidates were generated with Unsloth's Qwen3.8 27B GGUF in BF16 (Qwen3.8-27B-BF16, served as qwen3.8-27b-bf16) through llama.cpp with a 32,768-token context and the qwen model preset. The current workflow uses Ghidra's decompilation as a mechanical seed, asks Qwen only for a meaningful contract and bounded source edits, follows valid edits as a repair trajectory, and retains the best result measured by binary-comp. Gemma 4 31B IT BF16 was also benchmarked, but none of its candidates is retained in the source tree.

The scores below were remeasured from the current source tree with MSVC 4.1 and binary-comp on 2026-08-16. Logged time includes managed-server startup, generation or repair, build, and comparison for the retained candidate. It excludes unsuccessful exploratory runs and repair turns made after a resumable candidate had already been produced; those remain available in the run logs.

Address Function Current similarity Logged time
0x0040215C IsLevelIndexInRange 65.31% 3m 06.4s
0x00402205 ShowLevelComplete 93.96% 5m 26.1s
0x0040250C SubmitHighScore 97.37% 1m 49.7s
0x004026D0 RenderScoreboard 95.40% 2m 30.9s
0x00402BF3 AddLevelToTable 73.85% 4m 04.6s
0x00402DC6 UpdateGameDisplay 78.18% 1m 36.4s
0x00402EE7 GetLevelDisplayInfo 71.70% 2m 15.9s
0x00402FD5 GetLevelIndex 87.27% 5m 48.3s
0x00403430 IsBombAtLevel 80.62% 27.3 s
0x00403607 PlaceLevel 80.00% 2m 32.6s
0x00404764 DrawLevelIndicator 97.33% 1m 35.6s
0x0040499F RenderCornerGlyph 69.83% 1m 50.9s
0x00404D6E DrawBitmapPair 98.67% 1m 04.6s
0x00404E3B RenderBitmapToWindow 94.74% 26.4 s
0x00404ECE DrawBitmapToWindow 95.59% 59.3 s
0x00404F8B DrawScore 96.20% 1m 32.4s
0x0040507F DrawScorePanel 97.97% 2m 45.0s
0x0040525D DrawScoreDigits 94.20% 2m 13.7s
0x00405343 RenderScoreOverlay 77.78% 1m 28.1s
0x0040546B DrawFrameBorder 84.93% 37.0 s
0x0040552E ExplodeBomb 64.71% 1m 50.9s
0x0040560E RenderExplosionWave 85.07% 5m 40.0s
0x0040591A DrawBombExplosion 89.31% 1m 45.8s
0x00405AF5 DrawPausedOverlay 96.21% 1m 05.1s
0x00405C52 HighScoreDialogProc 79.82% 41.5 s
0x00405DB8 ScorePanelDialogProc 50.34% 1m 53.1s
0x00405F72 DemoVersionDialogProc 46.31% 4m 18.5s
0x004061D3 OrderDialogProc 53.16% 10m 46.5s
0x00408854 DrawStartButton 70.80% 1m 35.6s
0x00408A51 UpdateMainMenuState 97.56% 28.9 s
0x00408AD3 InitializeGameLevel 83.72% 5m 35.7s
0x00408E6A LoadLevelData 77.50% 52.5 s
0x00408F02 LoadHighScores 79.61% 1m 15.0s
0x00409092 SaveHighScores 94.12% 1m 25.9s
0x0040910C LoadLevelBitmaps 98.93% 7m 08.4s
0x00409DB6 IsRatsHelpFile 91.67% 41.5 s

Reconstruction attempts for the following addresses were exhausted without retaining a source implementation. Time spent is cumulative across all logged attempts.

Deferred address Time spent Outcome
0x00401000 10m 29.7s Function too large for a bounded first pass; original 1.13% scaffold retained
0x00403840 13m 10.0s No compilable candidate; too large for a fast bounded pass
0x00406674 Skipped: too large for a fast bounded pass

Acknowledgements

Special thanks to: