> Generated by `tools/backlog.py render` from `.run/backlog.jsonl`. These are functions the Phase-21 automation got **close** on but did NOT byte-match. The whole-binary byte-gate is the sole arbiter (G3/P9): **byte-matches bank and are NOT listed here** — only genuine near-misses/blockers are. Ranked by hand-session priority: **reach** (×N propagation leverage) → **closeness** (match_one mismatch count, lower = closer) → **size**. Each row's `best_draft` is the closest C the machine reached — resume from there.
**Open near-misses:**495 · by status {'near': 416, 'failed': 79} · by class {'WAVE': 20, 'plumbing': 42, 'other': 23, 'STUB': 1, 'schedule': 61, 'loose-typing': 7, 'struct': 60, 'regalloc-order': 58, 'remat': 4, 'iv-combine': 3, 'GIANT': 8, 'STRUCT': 7, None: 200, 'PINS': 1}
**Open near-misses:**644 · by status {'near': 561, 'failed': 83} · by class {'WAVE': 21, 'plumbing': 42, 'other': 23, 'STUB': 1, 'schedule': 61, 'loose-typing': 7, 'struct': 60, 'regalloc-order': 58, 'remat': 4, 'iv-combine': 3, 'GIANT': 8, 'STRUCT': 7, None: 348, 'PINS': 1}
| # | addr | reach | class | nins | status | closeness | where it stuck | best draft |
| 232 | func_8015CA28 | 1 | schedule | 91 | near | 0 | none — MATCH (goto forces the func_80161240 block to the function tail; bnez-to-end layout vs early-return fall-through) | `.run/backlog_drafts/func_8015CA28.c` |
| 233 | func_801820DC | 1 | struct | 94 | near | 0 | none — MATCH (switch on entity state at 0x34, inner jtbl on (s8)(u70>>8)) | `.run/backlog_drafts/func_801820DC.c` |
| 346 | func_8017F290 | 1 | regalloc-order | 86 | near | 2 | none — MATCH. Two levers: (1) hold &D_801270CC in a `int *state` local so its | `.run/backlog_drafts/func_8017F290.c` |
| 357 | func_80171B4C | 1 | schedule | 70 | near | 3 | 3-off in the tail only (body+prologue MATCH via $s1 pin on arg1). gcc fills the | `.run/backlog_drafts/func_80171B4C.c` |
| 364 | func_80140E6C | 1 | schedule | 37 | near | 4 | 4 ins — each save-across-call copy (move s1,v0 / move s0,v0) should fill the NEXT jal's delay slot (target) but gcc-2.7.2 sched ties the copy with the next call's arg-setup at priority 2 and the LUID tie-break (rank_for_schedule) keeps the copy first, so reorg fills the slot with the arg-setup instead; no C reshape found that flips the LUID/priority order without breaking the OR-chain regalloc. | `.run/backlog_drafts/func_80140E6C.c` |
| 432 | func_8016E9EC | 1 | schedule | 53 | near | 11 | 11 left — all GNU scheduler/canon tie-breaks (regs all match via pins): (a) prologue hoists `addiu a1,0x1C` into the save block; (b) first lbu reads $a0 not $s1 (incoming-arg still live); (c) iCopy copy `addu s4,s0` lands early (scheduler) vs target's func_800D2CA8 delay-slot; (d) `addu a2,s2,s3` vs target `s3,s2` commutative-canon (unflippable w/o breaking load order). Permuter can't run (register __asm__ pins rejected, cookbook §5a). | `.run/backlog_drafts/func_8016E9EC.c` |
| 446 | func_80161CD0 | 1 | regalloc-order | 20 | near | 14 | param_2 must survive the call in $s0; try plain C first then pin to $16 | `.run/backlog_drafts/func_80161CD0.c` |
| 460 | func_80148E54 | 1 | regalloc-order | 67 | near | 40 | 67/67 ins, structure+branches+block-order+body-reload all match; residual is pure register-allocation order (permuter-eligible, no register-pins). (1) masked first-angle folds andi->$s1; target keeps it in $a0 then copies addu $s1,$a0 lazily into the beq delay-slot (the s1-copy schedules 1 slot later). (2) the 0x80 const is CSE-held in $a2 across both test+body; target re-materializes 0x80 in $v0 per block (it gets clobbered by `srl v0,v1,8`), which both renames the compare regs ($v0 not $a2) AND fills the body's reload load-delay slot (mine emits a nop there). (3) single-exit ret funnels $v1->$v0 (move v0,v1) where target writes $v0 directly. All three are gcc regalloc/value-prop tie-breaks no C reshape steered (tried: $v0 pin, temp-split, low-mask-reuse compare, memory-clobber-vs-volatile reload) — leave for the permuter. | `.run/backlog_drafts/func_80148E54.c` |
| 461 | func_8017B0E4 | 1 | other | 61 | near | 51 | none — MATCH (expected; clean six-call sign-extend-store pattern) | `.run/backlog_drafts/func_8017B0E4.c` |
| 462 | func_80166F58 | 1 | schedule | 69 | near | 55 | regs/loop/logic/%lo-fold all match (pinned $s0-$s6); residual is reorg.c delay-slot fill — target replicates `addiu $v0,$s2,1` into the 3 skip-branch delay slots + commits `addu $s2,$v0,0` at merge (my codegen emits in-place `addiu $s2,$s2,1` w/ nop slots, -1 ins); plus prologue short-arg promote-then-move not fusing (`sra $a1;addu $s4,$a1,0` vs my fused `sra $s4,$a1`). Not source-steerable (pins block permuter). | `.run/backlog_drafts/func_80166F58.c` |
| 463 | func_80177DA8 | 1 | remat | 63 | near | 58 | gcc narrows `&0xfffffeff` on a 16-bit (lhu) value to `andi 0xfeff` (inline), but target hoisted -0x101 into $t4 (and-reg) in the loop preheader — a gcc-2.7.2 LICM-before-combine pass-ordering quirk unsteerable from C; that 1-ins gap cascades the constant-reg numbering + tail reassoc. Secondary: single-IV loop ptr anchors at +0xA (short, 2 accesses) not +0xC (byte) like target. | `.run/backlog_drafts/func_80177DA8.c` |
| 464 | func_80141A60 | 1 | other | 76 | near | 63 | none — MATCH (76 ins, relocation-masked). Body byte-identical; target reserves an 8-byte | `.run/backlog_drafts/func_80141A60.c` |
| 465 | func_801457A4 | 1 | other | 79 | near | 74 | MATCH at -O0 (79 ins, reloc-masked, real -O0 flags). BANKING BLOCKER: this -O0 fn lives in the -O2 main ov_SC01_077.c; needs its own §18 -O0 split at 0x801457A4 (existing _o0.c covers 0x8013B568..0x8013C98C only). match_one is -O2 -> WRONG here. | `.run/backlog_drafts/func_801457A4.c` |
| 466 | func_8016B6BC | 1 | schedule | 94 | near | 80 | 12-off, all in the last 12 ins (tail). First 82 ins byte-match. Target materializes | `.run/backlog_drafts/func_8016B6BC.c` |
| 471 | func_8015E018 | 1 | loose-typing | 47 | failed | | none — MATCH (signed char forces lbu+sll24/sra24 sign-extend; default char is unsigned in this toolchain) | `.run/backlog_drafts/func_8015E018.c` |
| 472 | func_80159A20 | 1 | struct | 58 | failed | | none — MATCH. Key lever: block tail-MERGE (gcc shared `bne ...,$L1` cross-jump) by making the two return-tests STRUCTURALLY DIFFERENT — positive `if(==){goto mask;} return;` per branch (not `if(!=) return;`), so neither tail can merge. Then fix block ORDER by inverting the outer test (`if(d!=1)` makes the ==4/return path the inline fall-through, the d==1/p[2] path the forward `beq`-target block — matching the target layout). | `.run/backlog_drafts/func_80159A20.c` |
| 473 | func_8013E958 | 1 | schedule | 63 | failed | | 2 ins swapped — cc1 emits `andi 0x7f; andi 0xff; beqz`(tests doubly-masked) but target wants `andi 0x7f; beqz; andi 0xff`(andi 0xff in the first beqz delay slot). Loop fully byte-matches; struct/ptr fold idioms nailed (D_80115110.q[0x16] hoists base+0x58 disp, Cell D_80115188[i].v folds %lo). func_800D0488 takes (m&0xFF) arg. s16 m -> 2 mism (right ins count); s32 m fixes order but merges andi+arg-move (62 ins, 1 short). | `.run/backlog_drafts/func_8013E958.c` |
| 474 | func_8013E83C | 1 | plumbing | 71 | failed | | none — MATCH (scalar global stores + 2 conditional calls; §3-T4 branch-polarity invert on the &0xFF test) | `.run/backlog_drafts/func_8013E83C.c` |
| 475 | func_8017B368 | 1 | schedule | 74 | failed | | dead table-path off by 1 ins — target keeps src*16 live in $s0 & computes 2nd arg src16+(base+8) in call1 delay slot; every C form either folds to e+8 (73 ins) or CSEs base into an extra saved reg $s2 (75 ins). All live code (prologue, sltiu dispatch, copy path, 4 struct-copy stores, tail) matches; only the unreachable func_8012F214 path scheduling diverges. | `.run/backlog_drafts/func_8017B368.c` |
| 476 | func_80180F10 | 1 | schedule | 75 | failed | | none — MATCH (aggregate-initializer form schedules const setup after the prologue saves) | `.run/backlog_drafts/func_80180F10.c` |
| 477 | func_8016C188 | 1 | schedule | 79 | failed | | none — MATCH (79 ins, match_one). Levers: invert if to if(iVar1!=0){big}else{small} | `.run/backlog_drafts/func_8016C188.c` |
| 479 | func_801418F8 | 1 | loose-typing | 90 | failed | | none — MATCH (90 ins). Keys: D_8011511A is `volatile u16` (store-2-then-read must NOT const-fold), read ONCE into a `u16` local t (an `unsigned int` local adds a stray andi 0xffff; reusing t feeds the single lhu to both sltiu and t-3), and the lone D_80115158 store after the if yields the delay-slot-fill-from-target dup of `addiu 0x106`. | `.run/backlog_drafts/func_801418F8.c` |
| 483 | func_801506A4 | 1 | loose-typing | 95 | failed | | none — MATCH (95 ins). switch var must be int not u16 (u16 adds andi 0xffff promotion mask) | `.run/backlog_drafts/func_801506A4.c` |
| 484 | func_8016BFD0 | 1 | struct | 95 | failed | | none — MATCH (95 ins). memcpy(d,s,4) -> lwl/lwr/swl/swr; MATRIX work buf as s32[16] (0x40, fills 0x10-0x4f, t[] = buf[5..7]); two SVECTOR locals at 0x50/0x58; natural saved-reg order s0=param_5/s1=param_4/s2=iVar5 | `.run/backlog_drafts/func_8016BFD0.c` |
| 485 | func_801596F0 | 1 | struct | 97 | failed | | none — MATCH (97 ins). Loop 1 = struct-ptr for-loop (E38, flag@+4) -> -4 guard / +4 test via gcc biv-elim. Loop 2 needed an EXPLICIT if-guard + do-while with the FIELD pointer as the loop var (so the giv init emits FIRST in the preheader, before the &D invariant block), a non-volatile re-tie barrier on d=&D_800AFAE8 (keeps &D BARE so +4/+0x88 add off it = +1 ins vs the folded %lo, matching target), an explicit u32 mask hoisted first (fills the guard's delay slot), and a separate range check fp<d+4 // fp>=d+0x88. | `.run/backlog_drafts/func_801596F0.c` |
| 486 | func_8017B490 | 1 | schedule | 97 | failed | | none — MATCH (97 ins). Keys: (1) inverted if so the SV4 block-copy else is fallthrough and block A is bnez-far (cond = (u32)&D_801DA73C >= 0xB, a relocated-symbol compare gcc can't fold); (2) SV4{s16 a,b,c,d} align-2 8-byte copy -> lwl/lwr/swl/swr, locals at sp+0x10/0x18; (3) the DEAD block-A index = (s32)&D_801DA73C into D_8018A45C[] (WAVE16=two SV4) needs BOTH source addrs (&[idx].v0,&[idx].v1) in EXPLICIT TEMPS before call1 so base (&D_8018A45C) dies pre-call -> temp $v0 (not callee-saved), forcing the natural 2-reg alloc idx=$s0/param_1=$s1 + src2 into call1's delay slot; (4) tail: D_801DA794/D_801DA78C are SV4, re-read .a/.b/.c via lh sign-extended to s32 stores. | `.run/backlog_drafts/func_8017B490.c` |
| 487 | func_80182C9C | 1 | schedule | 101 | failed | | none — MATCH (101 ins). Inner-switch break-to-shared-tail must be written as | `.run/backlog_drafts/func_80182C9C.c` |
| 467 | func_8017F290 | 1 | regalloc-order | 86 | near | 2 | none — MATCH. Two levers: (1) hold &D_801270CC in a `int *state` local so its | `.run/backlog_drafts/func_8017F290.c` |
| 480 | func_80171B4C | 1 | schedule | 70 | near | 3 | 3-off in the tail only (body+prologue MATCH via $s1 pin on arg1). gcc fills the | `.run/backlog_drafts/func_80171B4C.c` |
| 487 | func_80140E6C | 1 | schedule | 37 | near | 4 | 4 ins — each save-across-call copy (move s1,v0 / move s0,v0) should fill the NEXT jal's delay slot (target) but gcc-2.7.2 sched ties the copy with the next call's arg-setup at priority 2 and the LUID tie-break (rank_for_schedule) keeps the copy first, so reorg fills the slot with the arg-setup instead; no C reshape found that flips the LUID/priority order without breaking the OR-chain regalloc. | `.run/backlog_drafts/func_80140E6C.c` |
| 570 | func_8016E9EC | 1 | schedule | 53 | near | 11 | 11 left — all GNU scheduler/canon tie-breaks (regs all match via pins): (a) prologue hoists `addiu a1,0x1C` into the save block; (b) first lbu reads $a0 not $s1 (incoming-arg still live); (c) iCopy copy `addu s4,s0` lands early (scheduler) vs target's func_800D2CA8 delay-slot; (d) `addu a2,s2,s3` vs target `s3,s2` commutative-canon (unflippable w/o breaking load order). Permuter can't run (register __asm__ pins rejected, cookbook §5a). | `.run/backlog_drafts/func_8016E9EC.c` |
| 589 | func_80161CD0 | 1 | regalloc-order | 20 | near | 14 | param_2 must survive the call in $s0; try plain C first then pin to $16 | `.run/backlog_drafts/func_80161CD0.c` |
| 606 | func_80148E54 | 1 | regalloc-order | 67 | near | 40 | 67/67 ins, structure+branches+block-order+body-reload all match; residual is pure register-allocation order (permuter-eligible, no register-pins). (1) masked first-angle folds andi->$s1; target keeps it in $a0 then copies addu $s1,$a0 lazily into the beq delay-slot (the s1-copy schedules 1 slot later). (2) the 0x80 const is CSE-held in $a2 across both test+body; target re-materializes 0x80 in $v0 per block (it gets clobbered by `srl v0,v1,8`), which both renames the compare regs ($v0 not $a2) AND fills the body's reload load-delay slot (mine emits a nop there). (3) single-exit ret funnels $v1->$v0 (move v0,v1) where target writes $v0 directly. All three are gcc regalloc/value-prop tie-breaks no C reshape steered (tried: $v0 pin, temp-split, low-mask-reuse compare, memory-clobber-vs-volatile reload) — leave for the permuter. | `.run/backlog_drafts/func_80148E54.c` |
| 607 | func_8017B0E4 | 1 | other | 61 | near | 51 | none — MATCH (expected; clean six-call sign-extend-store pattern) | `.run/backlog_drafts/func_8017B0E4.c` |
| 608 | func_80166F58 | 1 | schedule | 69 | near | 55 | regs/loop/logic/%lo-fold all match (pinned $s0-$s6); residual is reorg.c delay-slot fill — target replicates `addiu $v0,$s2,1` into the 3 skip-branch delay slots + commits `addu $s2,$v0,0` at merge (my codegen emits in-place `addiu $s2,$s2,1` w/ nop slots, -1 ins); plus prologue short-arg promote-then-move not fusing (`sra $a1;addu $s4,$a1,0` vs my fused `sra $s4,$a1`). Not source-steerable (pins block permuter). | `.run/backlog_drafts/func_80166F58.c` |
| 609 | func_80177DA8 | 1 | remat | 63 | near | 58 | gcc narrows `&0xfffffeff` on a 16-bit (lhu) value to `andi 0xfeff` (inline), but target hoisted -0x101 into $t4 (and-reg) in the loop preheader — a gcc-2.7.2 LICM-before-combine pass-ordering quirk unsteerable from C; that 1-ins gap cascades the constant-reg numbering + tail reassoc. Secondary: single-IV loop ptr anchors at +0xA (short, 2 accesses) not +0xC (byte) like target. | `.run/backlog_drafts/func_80177DA8.c` |
| 610 | func_80141A60 | 1 | other | 76 | near | 63 | none — MATCH (76 ins, relocation-masked). Body byte-identical; target reserves an 8-byte | `.run/backlog_drafts/func_80141A60.c` |
| 611 | func_801457A4 | 1 | other | 79 | near | 74 | MATCH at -O0 (79 ins, reloc-masked, real -O0 flags). BANKING BLOCKER: this -O0 fn lives in the -O2 main ov_SC01_077.c; needs its own §18 -O0 split at 0x801457A4 (existing _o0.c covers 0x8013B568..0x8013C98C only). match_one is -O2 -> WRONG here. | `.run/backlog_drafts/func_801457A4.c` |
| 612 | func_8016B6BC | 1 | schedule | 94 | near | 80 | 12-off, all in the last 12 ins (tail). First 82 ins byte-match. Target materializes | `.run/backlog_drafts/func_8016B6BC.c` |
| 616 | func_8015E018 | 1 | loose-typing | 47 | failed | | none — MATCH (signed char forces lbu+sll24/sra24 sign-extend; default char is unsigned in this toolchain) | `.run/backlog_drafts/func_8015E018.c` |
| 617 | func_80159A20 | 1 | struct | 58 | failed | | none — MATCH. Key lever: block tail-MERGE (gcc shared `bne ...,$L1` cross-jump) by making the two return-tests STRUCTURALLY DIFFERENT — positive `if(==){goto mask;} return;` per branch (not `if(!=) return;`), so neither tail can merge. Then fix block ORDER by inverting the outer test (`if(d!=1)` makes the ==4/return path the inline fall-through, the d==1/p[2] path the forward `beq`-target block — matching the target layout). | `.run/backlog_drafts/func_80159A20.c` |
| 618 | func_8013E958 | 1 | schedule | 63 | failed | | 2 ins swapped — cc1 emits `andi 0x7f; andi 0xff; beqz`(tests doubly-masked) but target wants `andi 0x7f; beqz; andi 0xff`(andi 0xff in the first beqz delay slot). Loop fully byte-matches; struct/ptr fold idioms nailed (D_80115110.q[0x16] hoists base+0x58 disp, Cell D_80115188[i].v folds %lo). func_800D0488 takes (m&0xFF) arg. s16 m -> 2 mism (right ins count); s32 m fixes order but merges andi+arg-move (62 ins, 1 short). | `.run/backlog_drafts/func_8013E958.c` |
| 619 | func_8013E83C | 1 | plumbing | 71 | failed | | none — MATCH (scalar global stores + 2 conditional calls; §3-T4 branch-polarity invert on the &0xFF test) | `.run/backlog_drafts/func_8013E83C.c` |
| 620 | func_8017B368 | 1 | schedule | 74 | failed | | dead table-path off by 1 ins — target keeps src*16 live in $s0 & computes 2nd arg src16+(base+8) in call1 delay slot; every C form either folds to e+8 (73 ins) or CSEs base into an extra saved reg $s2 (75 ins). All live code (prologue, sltiu dispatch, copy path, 4 struct-copy stores, tail) matches; only the unreachable func_8012F214 path scheduling diverges. | `.run/backlog_drafts/func_8017B368.c` |
| 621 | func_80180F10 | 1 | schedule | 75 | failed | | none — MATCH (aggregate-initializer form schedules const setup after the prologue saves) | `.run/backlog_drafts/func_80180F10.c` |
| 622 | func_8016C188 | 1 | schedule | 79 | failed | | none — MATCH (79 ins, match_one). Levers: invert if to if(iVar1!=0){big}else{small} | `.run/backlog_drafts/func_8016C188.c` |
| 624 | func_801418F8 | 1 | loose-typing | 90 | failed | | none — MATCH (90 ins). Keys: D_8011511A is `volatile u16` (store-2-then-read must NOT const-fold), read ONCE into a `u16` local t (an `unsigned int` local adds a stray andi 0xffff; reusing t feeds the single lhu to both sltiu and t-3), and the lone D_80115158 store after the if yields the delay-slot-fill-from-target dup of `addiu 0x106`. | `.run/backlog_drafts/func_801418F8.c` |
| 628 | func_801506A4 | 1 | loose-typing | 95 | failed | | none — MATCH (95 ins). switch var must be int not u16 (u16 adds andi 0xffff promotion mask) | `.run/backlog_drafts/func_801506A4.c` |
| 629 | func_8016BFD0 | 1 | struct | 95 | failed | | none — MATCH (95 ins). memcpy(d,s,4) -> lwl/lwr/swl/swr; MATRIX work buf as s32[16] (0x40, fills 0x10-0x4f, t[] = buf[5..7]); two SVECTOR locals at 0x50/0x58; natural saved-reg order s0=param_5/s1=param_4/s2=iVar5 | `.run/backlog_drafts/func_8016BFD0.c` |
| 630 | func_801596F0 | 1 | struct | 97 | failed | | none — MATCH (97 ins). Loop 1 = struct-ptr for-loop (E38, flag@+4) -> -4 guard / +4 test via gcc biv-elim. Loop 2 needed an EXPLICIT if-guard + do-while with the FIELD pointer as the loop var (so the giv init emits FIRST in the preheader, before the &D invariant block), a non-volatile re-tie barrier on d=&D_800AFAE8 (keeps &D BARE so +4/+0x88 add off it = +1 ins vs the folded %lo, matching target), an explicit u32 mask hoisted first (fills the guard's delay slot), and a separate range check fp<d+4 // fp>=d+0x88. | `.run/backlog_drafts/func_801596F0.c` |
| 631 | func_8017B490 | 1 | schedule | 97 | failed | | none — MATCH (97 ins). Keys: (1) inverted if so the SV4 block-copy else is fallthrough and block A is bnez-far (cond = (u32)&D_801DA73C >= 0xB, a relocated-symbol compare gcc can't fold); (2) SV4{s16 a,b,c,d} align-2 8-byte copy -> lwl/lwr/swl/swr, locals at sp+0x10/0x18; (3) the DEAD block-A index = (s32)&D_801DA73C into D_8018A45C[] (WAVE16=two SV4) needs BOTH source addrs (&[idx].v0,&[idx].v1) in EXPLICIT TEMPS before call1 so base (&D_8018A45C) dies pre-call -> temp $v0 (not callee-saved), forcing the natural 2-reg alloc idx=$s0/param_1=$s1 + src2 into call1's delay slot; (4) tail: D_801DA794/D_801DA78C are SV4, re-read .a/.b/.c via lh sign-extended to s32 stores. | `.run/backlog_drafts/func_8017B490.c` |
| 632 | func_80182C9C | 1 | schedule | 101 | failed | | none — MATCH (101 ins). Inner-switch break-to-shared-tail must be written as | `.run/backlog_drafts/func_80182C9C.c` |
# Community matching-model plan — publishing the LLM decomp drafter (parked / future use)
> **Status:** design note, not active work. Captured 2026-07-01 during the Phase-23 LLM-tier build, while
> the reasoning is fresh (R30). Belongs to the eventual **public flip** (Gen2 exit / Gen3) — do not start
> until BFM is presentable and the repo is public-ready. Sibling docs: `docs/gen2-mips-matching-model.md`
> (how we built BFM's specialist), `docs/matching-cookbook.md` (the idioms the model encodes),
> `PROJECT_CONTEXT.md` (the public-flip + licensing decisions).
## TL;DR
We built the first *working* LLM **matching-drafter** for PS1/PsyQ decomp: a LoRA fine-tune of
Qwen2.5-Coder-7B that drafts byte-exact C for `gcc-2.7.2-psx -O2 -G0 + maspsx 2.56`, gated by an
incorruptible whole-binary byte-gate (the model only affects *throughput*, never correctness). Three
things are publishable, in increasing order of community value:
1.**The weights (`bfm-match-7b-v*`)** — directly useful to anyone on the *same* toolchain; a rough
structural drafter for anyone else.
2.**The pipeline + a toolchain-agnostic training plan** — lets *any* decomp project train its own
specialist on its own pairs in an afternoon. Higher leverage than the weights.
3.**A universal, toolchain-conditioned model** (shared base + per-toolchain LoRA) trained on a *pooled*
multi-decomp corpus — the ambitious, landmark contribution. Feasible; decomp.me is the ready dataset.
The byte-gate is the load-bearing idea that makes all of this safe to share: a weak or wrong draft from
*any* model can never land a wrong match — it just reverts. So a "pretty good across many compilers"
model is a legitimate accelerator everywhere, with zero correctness risk.
## Why a matching model is compiler-specific (and what that means for sharing)
A matching drafter learns one compiler's **codegen fingerprint** — its register allocation, instruction
scheduling, reload/CSE/peephole choices, and the C idioms that trigger specific byte sequences (cookbook
§17–§28). Change the compiler and those bytes change:
| Audience | Transfer from BFM's v3 | Why |
|---|---|---|
| **Same toolchain** (gcc-2.7.2-psx, similar flags) | **Direct, byte-level** | Identical codegen. Real audience: the Phase-21 sweep byte-verified **4 gcc-2.7.2-psx twins** — Xenogears, Vagrant Story (CC0), Tomba, + the psx lineage `cc1`. |
| **Same gcc family, different flags/version** (2.6.x, 2.8, `-O1`/`-O3`/`-G8`) | **Partial** | Most passes shared; regalloc/scheduling/peephole drift breaks byte-exactness but structure transfers. |
| **Different target, same era** (N64 MIPS via IDO/gcc, PS2, Saturn SH2) | **Structural only** | ISA and ABI differ; useful as a scaffold, not a byte-matcher. |
**State:** Phase 23 in progress (NOT a phase end). **v3 is the current model** — `bfm-match-7b-v3` (Qwen2.5-Coder-7B QLoRA on **corpus-v3**), adapter at `models/bfm-match-7b-v3` (v2 kept as fallback at `models/bfm-match-7b`). **LM Studio is EJECTED** — serve via **`tools/serve_local.py`** (Unsloth GPU, OpenAI endpoint), NOT LM Studio. The **8-hour autonomous run (2026-06-30)** built local serving + the prompt fix + corpus-v3 + v3 + a production batch → **fleet 63.82%** (+502 byte-identical, $0), 136/136 byte-clean, 27 commits this session (local — **Drew pushes**, R6). Pipeline validated end-to-end: a free local model banks the small/setter bulk, including the empty-leaf class v2 couldn't. Corpus `datasets/match_pairs/` + `.venv-train` gitignored. (Phase 22 close `PhaseEnd_Phase22.md` is committed `commit:0325`.)
**State:** Phase 23 in progress (NOT a phase end). **v3 is the current model** — `bfm-match-7b-v3` (Qwen2.5-Coder-7B QLoRA on **corpus-v3**), adapter at `models/bfm-match-7b-v3` (v2 kept as fallback at `models/bfm-match-7b`). **LM Studio is EJECTED** — serve via **`tools/serve_local.py`** (Unsloth GPU, OpenAI endpoint), NOT LM Studio. The **8-hour autonomous run (2026-06-30)** built local serving + the prompt fix + corpus-v3 + v3 + a production batch → **fleet 63.82%** (+502 byte-identical, $0), 136/136 byte-clean, 27 commits this session (local — **Drew pushes**, R6). Pipeline validated end-to-end: a free local model banks the small/setter bulk, including the empty-leaf class v2 couldn't. Corpus `datasets/match_pairs/` + `.venv-train` gitignored. (Phase 22 close `PhaseEnd_Phase22.md` is committed `commit:0325`.)**UPDATE 2026-07-01 (this session):** measured v3-on-reach≥2 (banks 6/15, the band is reachable) + built **`tools/inject_capped_externs.py`** (the macro-extern-injection lever) and freed **11 reach-134 propagation-capped fns ×134** → **fleet 64.16%** (+0.34%, $0), 136/136 byte-clean, 1687 groups. v3 still served via `serve_local.py`. See the 2026-07-01 progress-log entry for the full arc + the deferred type-walled-14 + the strategic fork.
**NEXT TASK — DECIDED (Drew, 2026-06-30): do A first, THEN B.** (Sequencing rationale: A is $0 + immediate + realizes the ×134 lever this run set up, and A's new banks enrich the corpus that B trains on — so A-then-B compounds.)
@@ -73,4 +73,5 @@ The **whole-binary byte-gate** (`gate_stage`/`harvest_verify`, G3/P9) is the sol
- 2026-06-30: **T7 DEBUGGED + FIXED.** 3 Explore scouts (tooling / run-evidence / corpus) + a direct code read (R14 — which resolved a flat contradiction between two scouts) found **two independent bugs** in `lora_grind`'s gate path: **(A)** `good_sha()` passed `"<sha> <name>"` vs harvest_verify's bare `sha1()` → 0 banks for ALL binaries incl. 077 (so 077's "0/12" was a bug artifact); **(B)** `src/asm/out` defaulted to ov_SC01_077 → non-077 drafts dropped at the 077 stub-filter, silently. Fixed `gate_stage.run_gate` (binary-agnostic resolution + bare-hash normalize + a loud negative-control guard) + `lora_grind.good_sha`; byte-neutral (check-all 136/136). ov_SC01_000 spot-run **banked 7/15 (47%) byte-identical** (@commit:0322) → reach-2 `func_8017CE24` propagated ×2. **ROI:** 6/7 reach-1 → broad rotation is high bank-rate / low fleet-% ROI; the fleet lever is **reach≥2 targeting** + corpus-v3. Backlog now correctly classified (4× close=1 = grinder fuel). NEXT: **T8 corpus-v3** (struct types) + **T9 reach≥2 selection** + concurrent grinder.
- 2026-06-30 (cont.): **T9 reach≥2 targeting built + measured.** Added `lora_grind --min-reach N` (lazy sig-based reach oracle == `dedup_propagate`, validated 0-mismatch/60 + the func_8017CE24=2 ground truth; `--min-reach 2` ranks high-reach-first, naturally restricts to overlays). Bounded reach≥2 mass-run: ov_SC01_000's 15 reach≥2 (shared) stubs banked **0/15** (vs the reach-1 spot-run's 7/15) — the model is **weakest on shared code** (corpus skipped the `DEFINE_func` bodies + it's the regalloc/schedule tail). But **5/15 are close≤3 reach-134 near-misses** (3× close=1 = func_8012E27C/BF4C/AD64) → high-value permuter fuel (×134 each). **FINDING: reach≥2 model-only ≠ a fleet lever; the lever is reach≥2-draft → grinder-close (×134)**, which needs `grinder.py`'s per-binary fix (same class as T7) + a backlog `binary` field. (A foreground mass-run hit the 10-min Bash cap mid-2nd-batch; tree recovered clean via `git checkout`, check-all 136/136.) Details: `docs/gen2-mips-matching-model.md` "T9 RESULT". NEXT: the grinder per-binary fix (realize the reach-134 ×134 fuel), then corpus-v3.
- 2026-06-30 (cont.): **Grinder per-binary fix (5-layer) — built + validated; reach≥2 fuel re-characterized.** The grinder/backlog were ov_SC01_077-hardcoded **5 layers deep** — fixed all: `gate_stage` records `binary`, `backlog.FIELDS` keeps it, `backlog.load_best`/`_open_stubs` fleet-aware (stuck-local fns surface via their overlay record), `p16_permute.setup` takes asm-subdir, `grinder` per-binary resolution + grouped gating + None-`nins` allow. Validated: the 3 fresh reach-134 close=1 ov_SC01_000 fns now surface + resolve + gate per-000. **FINDINGS (byte-evidenced):** (a) the reach≥2 close=1 fuel is **MODEL semantic-misses** (7B drafts empty `void f(void){}` for trivial `return 1`/setters), NOT permuter fuel — a corrected draft banks them (**+3 byte-identical** via the fixed gate, @commit:0326); (b) ×reach is **propagation-capped** (inline-matched in `ov_SC01_077_a.c` → dedup "nothing to propagate" → banked ×1). So the reach-134 ×134 payoff routes through **corpus-v3** (leaf variety) + the **dedup-collapse**, NOT the permuter. check-all 136/136 throughout. Details: `docs/gen2-mips-matching-model.md` (grinder fix). NEXT: corpus-v3 + dedup-collapse.
- 2026-07-01 (**A: reach≥2 measurement + the macro-extern-injection lever — fleet 63.82%→64.16%, $0 LLM**): served v3 (`serve_local.py`, ~15s). **(1) Measured v3 on reach≥2** (`lora_grind --min-reach 2`, bounded): banked **6/15 in ov_SC01_005** (+1 in 006) — **v3 banks the reach≥2 band**, vs v2-pre-fix's 0/15 (T9). But the banks are mostly low-reach (the 005≡006 pair = ×2) and re-bank inline per overlay → the fleet-% needs the COLLAPSE. **(2) Scoped the dedup-collapse:** the broad-overlay inline banks are mostly already-propagated (8-hr sweeps) or reach-1; the real ×134 fuel is **29 reach-134 fns matched INLINE in ov_SC01_077 but `dedup_propagate`-skipped as "not self-contained"** (the §19/20 propagation cap; the Phase-20 backlog list). Diagnosed (R14): they fail `compiles_standalone` on **file-scope callee/data externs** (clean lever) or **local types** (fiddly). **(3) Built `tools/inject_capped_externs.py`** (the Phase-20-backlog "macro-extern-injection", never built): source each capped fn's externs from 077's OWN file-scope decls, inject block-scope → self-contained → liftable ×134. Byte-neutral (077 stays `d19c9580…`), fail-safe (only rewrites if then `compiles_standalone`). **Freed 8 (main .c) + 3 (`_a.c`) = 11 reach-134 fns**, propagated ×134, **9 new shared groups (1678→1687)**, fleet **63.82%→64.16%** (+0.34%, +1204 byte-identical), check-all **136/136**, dedup-check 0 failed, **$0 LLM**. cookbook **§28d** + SETUP inventory (R16/R21/R30). **(4) Type-walled subset (14 reach-134 fns) DEFERRED:** the §28b type-lift (`build_engine_types --strip`) is **collision-blocked** (`typedef vec differs between header and source`) — the inline-invented types (MATRIX/Buf/Vec3/Loc/vec/…) need manual name-collision reconciliation; low-ROI/fiddly, byte-gate+collision-guard correctly refused. The 14: `0x8014964C 0x8014A048 0x8014C8C8 0x801678F0 0x8016A8FC 0x801749C8 0x80177EA4` (main) + `0x8012A418 0x8012A464 0x8012C658 0x8012E778 0x8012EA90 0x80134310 0x801395D4` (_a). Blast radius of a lift = overlays ONLY (engine_types.h does NOT reach main/resident). **(5) Launched a bounded broad v3 harvest** (12 batches, propagate-every-4) for $0 corpus growth + incremental banks — running at write time. **NEXT:** verify+commit the harvest; then the strategic fork — (a) v3.1 flywheel retrain on the grown corpus (compounding, ~2h GPU), (b) the type-collision reconciliation (~+0.5%, fiddly), or (c) accept the automated ceiling and head to PhaseEnd. The injection lever now CLOSES the v3-reach≥2→×134 loop for future 077 banks (idempotent re-run).
- 2026-06-30 (**8-hour autonomous run, Drew away**): prompt-fix + local serving + corpus-v3 + v3 + big batch. **LM Studio ejected** → built **`tools/serve_local.py`** (Unsloth GPU serving as an OpenAI endpoint; the prebuilt llama-cpp-python CUDA wheel SIGILLs on this no-AVX-512 CPU, so the Unsloth/torch path is the reliable one). **PROMPT FIX** (`api_draft.LEAN_SYS` + `format_finetune.SYS`, synced): "translate EVERY instruction, never an empty body" — the v2 empty-leaf overfit, small-leaf band **0/3→2/3**, banked 3 on a fresh ov_SC01_001 batch. **CORPUS-V3**: `export_pairs` now mines the **1623 engine_core.h `DEFINE_func` macros** (the shared setters/return-const the model was blind to — 96.6% of v2 was overlay-unique) + `format_finetune` inlines `engine_types.h` structs → corpus **1312→2891**, trainable **2534+291** (2.5× v2). **v3 trained** (Qwen2.5-Coder-7B QLoRA, loss 1.275→0.085, ~2h), held-out gate-true eval **MATCH 23/40 (57.5%)** (generalizing; v2's mixed-set rate was lower). Grinder concurrent during training = 0 banks (permuter tail exhausted, Phase-22 reality). **Big batch** (v3, broad rotation, 25 binaries, propagate-every-3, **$0 LLM**): banked **~352 fns inline + 45 new shared groups** (1633→1678) → fleet **63.67% → 63.82%** (+502 byte-identical), **136/136 byte-clean**, 25 auto-commits. v3 repeatedly banked the **empty-leaf/setter class v2 couldn't** (func_8012E27C=return 1, 8012AD64/BF4C=setters). **Pipeline validated end-to-end:** a free local fine-tuned model harvests the small/setter bulk at $0, gated identically (G3/P9). Details: `docs/gen2-mips-matching-model.md` ("Corpus-v3 ... 8-hour autonomous run"). **NEXT (fresh session):** (1) **shared/reach≥2 targeting** (`lora_grind --min-reach 2` with v3 — so the setter banks propagate ×134 instead of re-banking inline per binary — the fleet-% lever); (2) **dedup-collapse** the per-binary inline setters; (3) **corpus-v4** = struct-giant types; raise `--max-nins` as the band lifts. Serve: `tools/serve_local.py --adapter models/bfm-match-7b-v3` (R21 / SETUP §Tooling inventory).
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Drew T