An external-model bake-off, not a banking session. Nothing was banked; that is
the next session's first task.
FINDINGS
- CARD FUEL is the biggest lever, bigger than model choice: the same 10 cards
went 4/10 -> 9/10 when seed_ref/tu_ref/decl_prior were injected. The
"60-instruction ceiling" was an artifact of withholding fuel.
- Sub-50 is near-free: 19/19 verified MATCH at $0.007/function, blind.
7,724 sub-50 open functions = 73.5% of the remaining set.
- A free model (stealth/ox-alpha) cracked a 611-ins function and a jtbl
exemplar, and distilled §206 — whose two negative results were
independently byte-confirmed before banking.
- §206 transfers WITHIN a family (40 turns -> 11) but NOT across (56 turns,
0 compiles). jtbl costs ~40 turns of learning per family, not per class.
TOOL FIXES (all negative-controlled)
- family_remap.gather_externs: took the first ALPHABETICAL extern across the
overlay's TUs, carrying two types swapped (sh/lh vs lbu/sb). Now prefers
the extracted unit's own file. Blocked a 4-member/2,444-ins family.
- atlas.member_lever: aprop_card was loaded and never read while a bare
ledger DIFF forced needs-autopsy. PURE now outranks it — rescues 32
members / 11 families / 3,810 ins.
- decl_prior._ASM_SYM: the %hi/%lo arm had never fired (0 of 1,210 over four
waves). jal 306->306 zero regressions, data 0->299.
- api_agent.py (new): --cards, --brief, --max-cost, nudge loop, 429
attribution + backoff, transport retry, non-fatal tool faults, and a
repeated-call guard.
RULES PROPOSED: R40 (exonerate the instrument before attributing a failure to
its subject — seven instances this session) and R41 (a cost/rate/yield number
ships with its denominator — I quoted $0.30 against a $6.31 bill).
- tools/glm_reconcile.py (NEW): aim GLM's reasoning at the DEF-side loose-typing wall (body + conflicting
TU decls + reconciliation toolkit -> consistent buildable byte-identical decls); captures reasoning
(.run/glm_reason/, idiom source R16); relax-in-any-TU-file + crash-robust call
- api_draft: REASON=1 saves the reasoning trace per draft (idiom mining on any GLM run)
- fix_arity_callers: --any-proto (relax any prototype, not just (void))
- RESULT: GLM's reasoning is expert-level (store-width/sh-vs-sw awareness, K&R promotion, independently
derives the cast idiom) but banks only 1/7 reconciliations; mechanical relaxation 0/7. The def-side
wall is INTRINSIC (narrow-param + byte-level addressing defeat reconciliation) — Fable5 §3c re-test
CONFIRMS the wall holds even vs a frontier reasoning model aimed directly at it. func_80175184 banked,
check-all 136/136
- MAXTOK env (default 512 = local v3 unchanged); reasoning models (GLM5.2) need a high cap or
they spend the budget on reasoning tokens and return empty content
- accumulate usage.cost from the response -> per-run $ + $/fn readout (OpenRouter reports it)
- the model rambled to 4096 tokens on no-stop-token drafts (~80-130s each); a
≤15-ins fn's C is ~100-300 toks, so 512 truncates the ramblers (gate-rejected anyway)
- CURRENT_PHASE: ≤15 band NOT saturated (~4,597 unique / 4,087 untried, SC03=12,497);
drafting ~15-18s/func serial is the throughput wall; architecture = phase-separate +
parallel-gate + vLLM batched serving (the real GPU lever); fresh bank-rate still TBD
LM Studio was ejected, so serving is ours now: tools/serve_local.py serves base+LoRA via
Unsloth (.venv-train cu128) as an OpenAI endpoint — no llama.cpp build (this CPU has no
AVX-512, which SIGILLs the prebuilt llama-cpp-python CUDA wheels). api_draft/lora_grind hit
it unchanged.
PROMPT FIX (api_draft.LEAN_SYS + format_finetune.SYS, kept in sync): 'translate EVERY
instruction, never an empty body' — a prompt test took the small-leaf band 0/3 -> 2/3 MATCH
(the v2 corpus overfit an empty void f(void){} leaf pattern). Validated end-to-end: a fresh
ov_SC01_001 batch banked 3 via the local server + better prompt.
CORPUS-V3 (export_pairs + format_finetune):
- export_pairs now ALSO mines the 1623 DEFINE_func macro bodies in engine_core.h (the shared
setters/return-const/dispatchers extract_defs never saw -> 96.6% of v2 was overlay-unique,
the root of the empty-leaf overfit). Corpus 1312 inline -> 2891 (1312 inline + 1579 macros).
- format_finetune inlines engine_types.h structs in the compile-filter so struct-using bodies
are KEPT not dropped: train 2534/2591 (97.8%) compile standalone (v2 was 1111 total).
The fine-tuned 7B (bfm-match-7b-v2) drafted real OPEN ov_SC01_077 stubs; whole-binary gate banked 4
(func_80160B34 func_8015CC74 func_8016084C func_801705C0). Sample: 9/22 match_one proxy -> 4/22
whole-binary banked (18%; the proxy->gate gap is the TU-plumbing wall). Model is format-robust (raw .s
== normalized). api_draft: NORMALIZE_ASM bridge (unused — model handles raw .s) + ghidra_c-empty fix.
Gives the no-tool local model the context the agents read: common.h, live matching cookbook
(COOKBOOK_FULL toggle), 2 byte-matched corpus examples. Qwen3.6-35B-A3B result: harness fixes
compile-fails but model stays stuck at fixed near-misses; FULL cookbook worse+2.3x slower than
curated (dilution, gate-confirmed). Bottleneck is model refinement, not context.
docs/gen2-mips-matching-model.md: the BFM/gcc-2.7.2 matching-specialist idea (LoRA on our own
gate-verified pairs — the corpus off-the-shelf RE LLMs lack). export_pairs.py mines 1307 banked
(asm<->C) pairs from build objects (asm/ is gitignored, so disasm the ROM-identical build, splat-like
format) + src defs -> datasets/match_pairs/{pairs,train,test}.jsonl (gitignored, 1174/133 split).
api_draft.py: TEMP env-tunable. .gitignore: datasets/ models/ weights.
Script equivalent of one worker_wave drafter for any OpenAI-compatible endpoint (LM Studio /
llama.cpp / vLLM / OpenRouter). Inlines asm+ghidra_c+toolkit, calls /chat/completions, extracts C,
iterates against match_one (diff fed back, keep best). Output scores as a 'local'/'glm' arm via
ab_score.py. Env: API_BASE/API_KEY/MODEL. Logic self-tested; HTTP is standard OpenAI format.
Drew T