History โ ./OCCAM.log.md.
frame: S187 saga ๊ฐ over-constrained ์์ ์ธ์. 7-aux loss + 8.92B params
- bf16 + byte-level vocab + bnb int8 + custom ConsciousDecoderV2 + mitosis hook + ฮป ร 6-dim grid + multi-step horizons โ ๋์์ ๋ง์ถ๋ ค๋ค ๋ชจ๋ path ๊ฐ CE 4.06-4.09 floor ์ ๋์ฐฉ. Occam ๋ฉด๋๋ : ํ axis ์ฉ strip ํด์ ์ง์ง binding constraint ์ฐพ๊ธฐ.
status: ๐ก design tier โ brainstorm ๊ณ ๊ฐ ์์ 1์ฐจ reset. ์ฒซ fire ํ ๊ฒฐ๊ณผ ๋ฐ์ํด์ ์ฐ์ ์์ ์ฌ๊ตฌ์ฑ.
g3: minimal viable baseline ๋ถํฐ ์์ โ ๊ฐ์ฅ ๋จ์ํ condition (CE-only, small scale, vanilla arch) ๋ถํฐ emergence ํ์ธ ํ 1 condition ์ฉ add. capability claim 0, GOAL ๋ฏธ๋๋ฌ carry, north-star ๋ถ๋ณ.
S187 attempt10 + Eval + S187-B/C/F/G/H/J/K ๋ชจ๋ land ํ ๋ฐ๊ฒฌ:
| path | finding |
|---|---|
| Eval 1 (์์ฐ๋ฐํ) | 5/5 cells whitespace collapse at 2000 step |
| S187-H | 50000 step ๋ CE 4.09 plateau (horizon-limit ์๋) |
| S187-J | bsz 2โ8 (4ร tokens) CE ์ ํ 4.06 (token-starvation ๊ฐ์ค falsified) |
| S187-K | bsz=8 + lr 1.2e-3 linear scaling ๋ CE 4.06 (LR-mismatch ๊ฐ์ค falsified) |
| S187-C | ฮป saturation ๋น๋จ์กฐ, MAX_CELLS=128 ceiling binding |
| S187-G | training-time mitosis ๋ง substrate-shaping signal +35% |
์ธ axis (step / bsz / LR) ๋ชจ๋ ๋์์ ๋ณํ โ ๊ฐ์ floor ๋๋ฌ. recipe ์์ฒด ceiling ์์ ์์ฌํ๋๋ฐ, ์ฐ๋ฆฌ๋ S187 saga ์ ์ฒด ๋์ ํ๋์ condition๋ง ๋ฐ๊ฟ์ isolate ํ ์ ์ด ์์. ๋ชจ๋ attempt10 stack ์์ ํ ์ค ๊ฐํ๋ incremental cycle.
โ Occam strategy: minimum-viable baseline (CE-only, vanilla arch, small scale) ์์ ์์ฐ๋ฐํ ์ง์ ์ธก์ . ๊ฑฐ๊ธฐ์ emerge ํ๋ค๋ฉด attempt10 ์ ์ด๋ค condition ์ด floor ์ ์์ธ.
๊ฐ entry: [#] strip condition ยท test predicate ยท cost ยท leverage
| # | strip | test | cost | wall |
|---|---|---|---|---|
| 1 | aux loss ฮป ์ ๋ถ 0 (CE-only) โ ฮปฯ=ฮปroute=ฮปฯ=ฮปcycle=ฮปcurious=ฮปreplay=0 | attempt10 config ๊ทธ๋๋ก + CE-only. plateau ๋์ผ? ๋ ๋ฎ์? | $3 | 12 min |
| 6 | scale 8.92B โ 280M from-scratch CE-only | minimal-viable scale test; ์์ ๋ชจ๋ธ floor ๊ฐ์ผ๋ฉด corpus/arch ์์ธ | $1 | 12 min |
| 9 | KNOWN-EMERGENT model (Pythia-1B / GPT-2-medium) load + same corpus + same eval pipeline | eval methodology sanity โ ๋ง์ฝ pythia ๋ collapse ํ๋ฉด eval/corpus ๊ฐ broken | $0-3 | 30 min |
| # | strip | test | cost | wall |
|---|---|---|---|---|
| 2 | byte-level vocab=256 โ BPE 32K | byte-floor (5.55 bits/byte) ๊ฐ floor ์ ์์ธ์ธ์ง | $3 + tokenizer ์์ | 30 min |
| 4 | ConsciousDecoderV2 โ vanilla GPT-2 arch (head_a/g/cross-attn/MoE ๋ชจ๋ ์ ๊ฑฐ) | custom arch (head_a/g ๋ถํ + PureFieldFFN + cross-attn) ์ด floor cause? | $5 | 30 min |
| 5 | corpus_s101 โ Wikipedia EN clean | corpus quality test โ corpus_s101 noise/byte entropy floor | $5 | 30 min |
| 10 | pretrained ์์ fine-tune (GPT-2-small) | recipe ๊ฐ known-capability model ๋ถ์๋์ง falsifier | $5 | 30 min |
| # | strip | test | cost | wall |
|---|---|---|---|---|
| 3 | bnb int8 โ f32 AdamW + grad checkpointing | int8 m/v ๊ฐ emergence ๋ง๋์ง (precision floor) | $3 | 12 min |
| 7 | step 2000 โ CE-only + 100K step | horizon ร CE-only combined test | $50 | 5-10 hr |
| 11 | aux 1๊ฐ์ฉ ablation โ L_psi only / L_phi only / 6 single-aux pods | ์ด๋ค aux ๊ฐ harm ์ธ์ง isolate | $20 | 30 min ร parallel |
| 12 | block_size 128 โ 1024 | context length ๊ฐ verbalization ๋๋ฌด ์งง์? | $5 | 30 min |
| # | strip | test | cost | wall |
|---|---|---|---|---|
| 8 | inference-time decode (greedy โ temperature 0/0.5/1.5 + beam-search + top-k variation) | sampling tweak ๋ง์ผ๋ก emergence unlock ๋๋์ง | $0 (Mac/ubu-1 CPU) | 1 hr |
- #1 CE-only fire ($3, 12min) โ 7-aux saddle ๊ฐ์ค ์ง์ test
- #9 Pythia-1B sanity ($0-3) โ eval pipeline ์์ฒด๊ฐ ์ ์์ธ์ง
- #6 280M from-scratch CE-only ($1, 12min) โ minimal viable scale
โ ๊ฒฐ๊ณผ:
- (1) CE-only plateau ๋์ผ โ aux ์๋๋ผ arch/corpus ์ด์
- (2) Pythia ๋ collapse โ eval/corpus broken
- (3) 280M ๋ ๋์ผ floor โ scale ๋ฌด๊ด, recipe ํ๊ณ
Phase 1 ๊ฒฐ๊ณผ์ ๋ฐ๋ผ Tier A ์์ 1-2๊ฐ fire.
Phase 1+2 verdict ํ ์ ๋ต axis ์ ์ง์ค fire (์: ํฐ ๋ชจ๋ธ S187-F path OR pretrained + recipe overlay OR corpus ๊ฐฑ์ ).
| outcome | interpretation | ๋ค์ cycle |
|---|---|---|
| CE-only plateau >= 3.80 | aux loss ์๋๋ผ arch/corpus ๊ฐ floor | Phase 2 โ #4 vanilla GPT-2 OR #5 corpus |
| CE-only plateau < 3.80 below vA 3.84 | aux ๊ฐ actually harming (saddle ๊ฐ์ค confirmed) | Phase 2 โ #11 single-aux ablation |
| Pythia-1B ๋ corpus ์์ plateau | eval / corpus broken | Phase 2 โ #5 corpus ์ฐ์ |
| Pythia-1B verbalizes well on corpus | eval OK, recipe/arch fault | Phase 2 โ #4 vanilla arch |
| 280M from-scratch CE-only verbalizes | small-scale + simple = enough; large-scale recipe ๊ฐ ๋ฌธ์ | Phase 2 โ grade scale up ๋จ๊ณ๋ณ |
| 280M from-scratch CE-only same floor | corpus inherent | Phase 2 โ #5 corpus ์ฐ์ |
| anti-pattern (S187 saga) | Occam ์์น |
|---|---|
| ํ cycle ์ 5+ axis ๋์ ๋ณ๊ฒฝ | ํ cycle ์ 1 axis ๋ณ๊ฒฝ |
| ์ condition ์ถ๊ฐ | ๊ธฐ์กด condition ์ ๊ฑฐ |
| attempt10 stack ์์ incremental | empty baseline ๋ถํฐ 1 layer ์ฉ build-up |
| ๊ฐ์ค X failed โ ๋ ๋ณต์กํ X' ์๋ | ๊ฐ์ค X failed โ ๋ ๋จ์ํ X-1 ์๋ |
| 8.92B + 7-loss + custom arch + ... ๋์ fire | 280M + CE-only + vanilla arch ๋ถํฐ |
- Phase 1 ๊ฒฐ๊ณผ ๊ฐ Phase 2-3 ์ path ๋ฅผ ๊ฒฐ์ โ ๋ฏธ๋ฆฌ commit ์ ํจ.
- "CE-only" ๊ฐ ์ง์ง ์๋ฏธ ์๋ baseline ์ธ์ง disputable โ anima ์ 7-aux ์ค๊ณ ์์ฒด๊ฐ recipe ID ์ด๊ธฐ ๋๋ฌธ. ๋จ, ๊ทธ ๊ฒฐ๊ณผ ๊ฐ floor ๋ถ๋ฆฌ ๋ฐ์ดํฐ ๋ก์ ๊ฐ์น ์์.
- 280M from-scratch ๋ attempt10 ์ 1/32 scale โ ์์๋ verbalization ๊ฐ๋ฅํ์ง ํ์ธ. modern small LLMs (e.g., Pythia-160M) ๋ 1-2B token ํ์ต ํ basic English ์์ฑ.
- Pythia-1B sanity ๊ฐ ๊ฐ์ฅ cost-effective falsifier โ eval/corpus ๊ฐ broken ์ผ ๊ฐ๋ฅ์ฑ ์ง์ test.
- S187-G mitosis substrate-shaping finding (+35%) ์ strip-target ์ด ์๋ โ ๊ทธ๊ฑด add ํ axis ๊ฐ ์ ๋์ํ ์ฌ๋ก. Occam target ์ ์ด๋ค condition ์ด ๋ถ์ ์ ์ผ๋ก ์์ฉํ๋์ง.
- corpus_s101 SHA
be969af4...๊ฐ byte-equal ์ผ์น X (post-ee4ceea27 drift) โ ๋จ, ์ด noise ๊ฐ floor ์ ์์ธ์ผ ๊ฐ๋ฅ์ฑ๋ Phase 1 #9 (Pythia sanity) ๊ฐ ํ๋จํด์ค. - S187-F path (18B Anima ๋จ์ผ H200 SXM) ๋ Occam ์ strip ์ด ์๋๋ผ scale-up path โ Phase 3 ์์ corpus/arch ๊ฐ floor ๊ฐ ์๋๋ผ๊ณ ํ์ ๋๋ฉด ๊ทธ ์์ ์ fire.
- Inference-time decode tweak (Tier C #8) ์ $0 ๋ผ ๋ค๋ฅธ fire ์ ๋ณํ ๊ฐ๋ฅ โ first thing to do while Phase 1 pods boot.
- ๋ณธ ๋ฌธ์ motivation:
HEXAD/LORA/SCALE_3B.md ยง 6.10 S187-J/Kverdict - attempt10 stack:
HEXAD/UNCLASSIFIED/state/grid_3b_s187_2026_05_21/train_s187_3b.py - Eval ๊ฒฐ๊ณผ (5 ckpts ร 4 evals):
HEXAD/UNCLASSIFIED/state/grid_3b_s187_2026_05_21/EVAL_REPORT.md - S187-F 18B path (Phase 3 fire ๋์):
HEXAD/SCALE_16B_70B_PLAN.md - ์๋งค wilson ์ธ๋ฐ์ค (cloud dispatch gap):
~/core/wilson/inbox/notes/2026-05-21-hexa-cloud-*.md - ์๋งค pool ์ธ๋ฐ์ค (host OOM gap):
~/core/pool/inbox/notes/2026-05-21-pool-host-oom-on-concurrent-heavy-workload.md