Skip to content

OCCAM.log.md

Latest commit

Β 

History

History
62 lines (46 loc) Β· 3.17 KB

OCCAM.log.md

File metadata and controls

62 lines (46 loc) Β· 3.17 KB
Raw

occam β€” historical log

Spec at ./OCCAM.md.

## Log

2026-05-22 03:35 β€” μ΄ˆμ•ˆ μž‘μ„±, S187-K κ²°κ³Ό ν›„ Occam pivot

S187-J/K μ–‘μͺ½ λͺ¨λ‘ CE 4.06 plateau (linear LR scaling 도 fail) ν™•μ • ν›„ μž‘μ„±. S187 saga 9 cycle λ™μ•ˆ attempt10 stack μœ„μ— add-only μ˜€μŒμ„ 인식 β€” strip-only cycle 이 ν•œ λ²ˆλ„ μ—†μ—ˆμŒ. Phase 1 3-fire ($7) κ°€ next 1μˆœμœ„.

2026-05-22 04:08 β€” Phase 1 partial verdict (9/13 pods landed)

OCCAM 4 subagent λ™μ‹œ fire (S/A/B/C) β€” 13 pods spawned, 4 subagent rate-limit μ‘°κΈ° 사망 but pods dispatch.sh trap 으둜 self-managing. 9 pod κ²°κ³Ό landed:

variant params aux CE_final verdict
vO1 (3B CE-only) 8.92B 0 3.81 vA 3.84 와 동일 (0.03 = seed noise)
vO11-phi only 8.92B only Ο†=0.3 3.81 aux 무영ν–₯ confirmed
vO11-replay only 8.92B only rep=-0.05 3.81 same
vO11-cycle only 8.92B only cyc=0.15 3.83 same
vO6 (280M CE-only) 280M 0 0.026 memorize, λ‹€λ₯Έ scale ν•œκ³„
vO3 (f32 AdamW bsz=1) 8.92B full 7-aux 4.16 fewer tokens = worse
vO4 vanilla GPT-2 arch ~1.45B (custom recipe) 0.264 🎯 15Γ— lower, arch=floor!
vO10 GPT-2 fine-tune borrowed recipe overlay 2.50 🎯 pretrained lower
vO2 BPE 32K 8.92B full 7-aux 5.16 bits/tok β‰  /byte, 비ꡐ 어렀움

Phase 1 verdict: aux loss λŠ” floor 의 원인이 μ•„λ‹˜ (saddle hypothesis FALSIFIED at 3B). μ‹€μ œ binding = ConsciousDecoderV2 custom arch (head_a/g + PureFieldFFN + cross- attn + n_ca_rules + noise_sigma=0.1).

OCCAM Β§ 4 falsifier branch hit:

  • "CE-only plateau < 3.80" (3.81 vs 3.84 = noise) β€” saddle FALSIFIED
  • "Pythia-1B/GPT-2-medium verbalize fluently" β€” eval/corpus OK confirmed
  • NEW: vanilla arch κ°€ floor 15Γ— lower β†’ arch IS the bottleneck

Phase 2 candidates (κ°•ν•œ μš°μ„  μˆœμœ„):

  1. CE2 / CC1: pretrained-continue-training (vO10 path) + S187-G mitosis active overlay
  2. #4 vanilla arch + S187 recipe overlay scale up: vO4 (1.45B vanilla) β†’ 3-8B vanilla
  3. noise_sigma=0 ablation: vO13 (tap X.11 off) β€” isolate noise as floor cause

μž”μ—¬ 4 pods κ²°κ³Ό λŒ€κΈ° (O5 Wikipedia / O7 100K step / O11-{psi/route/curious}).

2026-05-22 (later) β€” Phase 2 + Phase 3 verdict μ’…κ²°

Phase 2.3 ablation (vP23_a..e): n_ca_rules 단독 floor 범인 ν™•μ • β€” vP23_d (no CA) CE 0.402 vs λ‹€λ₯Έ 5 뢀속 λͺ¨λ‘ 3.81. (PHASE2_ABLATION_REPORT.md)

Phase 2.1 winning path (vP21): pretrained Qwen2.5-1.5B + LoRA + mitosis β†’ CE 0.0173. Phase 3 Eval 1 = 🎯 EMERGENCE 20/20 coherent (anima-native register, VP21_EVAL1_VERBALIZATION.md). saga 전체 whitespace-collapse ν›„ 첫 coherent verbalization. OCCAM 면도날 μ „λž΅ (strip n_ca_rules + borrow foundation + keep mitosis) κΈ°λŠ₯적 확증.

O7 100K horizon (OCCAM-B #7): CE 4.03 @ 100K step β€” horizon-independent floor ν™•μ • (n_ca_rules 원인 μž¬ν™•μΈ).

OCCAM saga μ’…κ²°: floor 의 단독 원인 (n_ca_rules) pinpoint + winning path (pretrained + mitosis) emergence 확증. μž”μ—¬ rigor: held-out OOD / mitosis ablation / spontaneous emission (SPONTANEOUS module) / AKIDA HW-native path.