Drew T
cee69c1bb5
Measured over 129 drafting agents in one session, per MATCHED instruction (the
only cost that matters, since a failed agent is billed in full):
sonnet 105 agents, 57 MATCH 4,289 tok/matched-ins (flat ~47% above 30 ins)
opus 24 agents, 11 MATCH 2,083 (m1 191-347: 1,291, 67%)
opus at 347-670: 1/9 7,158 <- the cliff, 2.92M tokens for ONE bank
fable escalation: 3/4 closed at ~1/3 the cost of the attempt it rescued
Sonnet's per-agent price was never the cost that mattered; cost per BANK is, and
it lost on that by 2.1x. The m2 wave should have been fable from the start.
Escalating SOONER is the standing finding — higher models crack harder functions
in fewer tokens. Tested twice now (S68 A/B, S69 measurement); do not re-derive a
cheap-tier argument from per-agent price a third time.
BFM-decomp
A matching decompilation of Brave Fencer Musashi (PlayStation, SLUS-00726, USA 1998) — the first public decompilation effort for this game.
What "matching" means
The goal is C source code that, compiled with the original-era toolchain (PsyQ 4.x / GCC 2.7.2-family + ASPSX via maspsx), produces a byte-for-byte identical SLUS_007.26 and, eventually, byte-identical overlay binaries. SHA1 checksums are the ground truth; "functionally equivalent" does not count.
No ROM content
This repository contains no game assets, no disassembly output, and no ROM-derived data — only source code, build configuration, symbol names/addresses, hashes, and documentation. To build or contribute you must provide your own dump of the game disc (4-track BIN/CUE, redump layout). See .gitignore for the firewall.
Project status
Latest (Phase 19, 2026-06-20): the project builds 136 binaries byte-identical from a clean tree (the EXE + the resident engine + all 134 location overlays);
make check-all→ 136/136. Fleet byte-identical-from-source is 58.0% (function-instance-weighted; see the PhaseEnds for the byte-weighted ~30% figure and what it includes). Shared engine functions are matched once inov_SC01_077and propagated ×134 viatools/dedup_propagate.py. (The narrative below is Phase-11/12-era; a full refresh is part of the public-flip prep.)
Gen1 (foundation) complete — the matching pipeline is proven end-to-end. make extract && make build && make check rebuilds SLUS_007.26 byte-for-byte identical (SHA1 143dbb89…) from C + assembly, reproducibly across many sessions.
- Compiler pinned by evidence:
gcc-2.7.2-psx -O2 -G0 -mips1 -mcpu=3000+maspsx --aspsx-version=2.56 --expand-div. - 52 functions hand-matched to byte-identical machine code — including the LZSS streaming decompressor — with a decomp-permuter + matching-cookbook "flywheel" to accelerate the next.
- 959 PsyQ SDK functions linked byte-identical (libcd, libgs, libgte, libspu/libsnd, libgpu, libc2, libmcrd, libapi/libcard, libetc) straight from the real PsyQ 4.0 libraries instead of re-decompiling them — bringing byte-identical-from-source coverage of the EXE to ~50%.
- File-loader / overlay system reverse-engineered, with the resident engine blob + location overlays' load addresses proven byte-identical against a live PCSX-Redux RAM dump.
About half the EXE is still INCLUDE_ASM stubs (correct bytes, not yet C), and the bulk of the game lives in compressed overlays inside the .CD archives — Gen2 (overlays & engine at scale) is underway:
- The build toolchain is binary-agnostic (one parameterized pipeline builds any binary), and the always-resident engine blob rebuilds byte-for-byte from source (SHA1
8e17e02f…) — the second binary reconstructed exactly, after the EXE — and is now 86% hand-matched C (123 / 146 functions, up from 0): its scripting turned out to be compiled-MIPS state/mode dispatch, not a bytecode VM, and the save-file + sound (SQV) formats are documented. The harvest used a reusable swarm-of-agents + bit-for-bit byte-gate method (a wrong match can't be accepted) —tools/harvest_verify.py+tools/match_one.py, which carry straight into the overlay phase. - A cross-binary deduplication pipeline is live: a Ghidra-free signer fingerprints all 134 location overlays, and the report finds ~9,000 byte-identical function groups shared across binaries (~28 MB of collapsible code) — a single engine function is byte-identical in all 134 overlays. This is "one match unlocks many": each engine match will be auto-credited across the overlay fleet.
Current phase and detailed progress live in phase-ends/ (newest PhaseEnd_*.md = current state); methodology, rules, and the full roadmap are in PROJECT_CONTEXT.md; environment setup in docs/SETUP.md.
This project is developed primarily by Claude Code driving Ghidra through an MCP server; see CLAUDE.md.
License
Private repository for now. AGPL-3.0 is planned at public release, modeled on sotn-decomp. tools/brave-CUE/ is CUE's BRAVE extractor (GPL, source included) and retains its own license.