Case study
Does reverse engineering need a bigger model? General and specialist models in Touhou reconstruction
An essay on REA’s site by N0zoM1z0 (13 October 2026) uses dated Touhou reconstruction commits to ask whether exact function matching needs a larger general model, a compiler-specialist model or better retrieval. It walks through TH09, TH095, TH04 and TH08 cases where later GPT-6-series sessions closed stubborn matches, and the TH08 audit that found accepted functions with wrong floating constants. Specialist and reinforcement-learning comparisons are proposed as future experiments; no trained specialist was evaluated.
- Targets
- Binary · Game
- Category
- Official REA: case studies
- Relationship
- Official REA case study
- Sources reviewed
Requirements & limits
- Prerequisites: The essay and its notes supply no game files, compilers or builds; each linked Touhou repository documents its own requirements.
- Scope: The essay is dated 13 October 2026, and its evidence notes record a review of repository history, maintained source and comparison notes that day without running game targets, compiling candidates or repeating historical runtime checks. Its examples are function-level comparisons recorded in linked TH09, TH095, TH04 and TH08 commits, not whole-game runtime checks.
- Limits: Later sessions inherited earlier source, diagnostics and failed experiments, so the before-and-after closures are not controlled estimates of model-size effects. The notes treat gpt-web as a changing label that does not identify a run’s model, per-run reasoning effort and budgets are not uniformly recorded, and the GPT-6-Luna-max closure is presented as a discovery across model tiers, not a measured smaller-model advantage.
- Testing evidence: The TH08 example comes from commit a393f400 (2 September), which target-checks every VC7 floating-literal relocation; the follow-up audits 1,548 relocations and corrects 12 stale rows in five accepted units. The essay’s concern that a reward built on the old check could reinforce wrong source is stated as an inference: the notes record no actual RL exploit or trained model doing so.
- License: Game rights stay with their owners.
- Attribution: The essay does not describe REA’s part in the cited sessions.
About this entry
The description was reviewed against primary sources. This does not establish runtime testing, security auditing or verified compatibility.
Primary sources
- website/evidence/reverse-engineering-models.md GitHub · morluto/rea · Revision e8da7e1
- website/public/blog/reverse-engineering-models/index.html GitHub · morluto/rea · Revision e8da7e1
- https://github.com/N0zoM1z0/th08/commit/a393f400786e4cc6cbb1000af0dad3276da12c78
More in Games
- LASRecomp: Link’s Awakening static recompilation · Case study
- ArkWeb: a two-game movement and collision experiment · Case study
- Perfect Dark: matching Nintendo 64 decompilation · Case study
- Super Mario Sunshine: a native PC port from an AI-assisted decompilation · Case study
- Majora’s Mask: matching US N64 decompilation · Case study
- DLSSNR-AMD: reimplementing DLSS 5 neural rendering for RDNA4 · Case study