Drew T
605a4b77c8
Drew asked whether waves cracked better before MAXTOK went 8000 -> 16000. Recording both halves of the answer so next session does not relitigate it from memory: CLEAN: raising to 16k did cause a real regression — draft completion 84-89% (8k) -> 41% on wave bt, 69% on bu — but the cause was a harness interaction, not the model. A 16k generation runs ~530 s at ~30 tok/s while STRAGGLER_GRACE was 120 s, so agents were cut off mid-thought with no draft. Grace at 700 s fixed it; completion has run 97-99% since. CONFOUNDED: on banks per draft the 8k era looks better (S59: 1,335 of 2,996 = 44.6%; today's best 16k waves dd 34.7%, de 29.7%) — but the populations differ completely. 8k waves had never-drafted work; today's draw from skeletons that refused six times. Budget and exhaustion moved together, so neither figure isolates the other. Neither should be cited as evidence about the budget. AGAINST the simple story: truncated-turn rate is INVERSELY correlated with bank rate (cx 8.7% trunc/43.9% bank, dd 8.3%/51.4% vs dl 1.3%/0.5%, ej 0.6%/0%). Budget exhaustion driving the decline would produce the opposite relationship. THE A/B: split ONE wave's card pool — half the shards at 8k, half at 16k, same generation mix, same binaries, same gate, same tree, grace 700 s in both arms. Compare banks per DRAFT and per GATE MINUTE. Holding the population constant is the whole point; every historical comparison fails exactly there. If 8k matches 16k, the cheaper budget also buys more agents per unit 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.