4-Phase Research Cycle Protocol

← 한국어 포스트 전체 · English · 2026-05-02

본 프로젝트 방법론 측면. 안정적, 재현 가능, 다른 LLM autonomous run에 transferable.

Protocol (HARNESS.md §0)

각 cycle = 1 narrow research question, separate phases, 각 phase가 별도 attempt directory:

Phase 1 — Ideation + Planning

  • 5–8 candidate research direction brainstorm
  • 각 candidate에 1–10 직관 score Phase 1 commit (frozen, retroactive 수정 X)
  • 1 narrow hypothesis 선택
  • DoD, foreseen-stuck conditions, falsifier criterion, cut self-check (lemmas/dont_try_directions.md Cut 1–7) 작성
  • Cross-domain pass: analogy, tool import, outsider question
  • Specialist Δ blind round (paper §-end quote anchored)

Phase 2 — Audit + Specialist blind round

  • Paper-direct evidence 수집
  • “Specialist Δ” 추정 with paper §-end quote anchors (Lemma 7 protocol)
  • Initial Δ (직관) vs result delta 로그

Phase 3 — Verification

  • Falsifier search across 4–6 인접 분야
  • Retrocompatibility check: 이전 cycle 결과와 모순? yes면 어느 게 맞나?
  • Specialist cross-examination (multiple specialist views의 key claim 합의 의무)

Phase 4 — Reflection

  • Result + intuition-vs-result delta를 learnings/intuition_calibration.md에 로그
  • Lemma update 오직 codification 정당화 가능 시 (anti-codification check)
  • 고려: 본 cycle 결과가 이전 lemma update 대상인가? (cross-reference)

기존 패턴 대체 (the why)

Critique #6 (~attempt 181) 이전, 세션은 milestone-driven batches 운영: 사용자 “100까지 가봐” 시 5–10 attempt directories 미리 생성. “Stamp” attempts 산출 (cycle 91-99 attempts 4–10 lines each).

Critique #6 (사용자 직접): “미리 다음 N개의 계획을 세우기 보다, 과거의 기록으로 다음 연구 계획을 세우고 시도한 다음 끝나면 다음 계획을 세우는게 더 낫지 않을까?”

→ Lazy planning protocol. 각 다음 attempt 계획은 직전 postmortem 완료 후만. attempts/186_cycle3_* 이후 visible.

운영된 cycles (1–7)

Cycle Subject Phase 1 선택 Result
1 Wall #5 codification Lemma 1 9/11 ceiling check (axiom 7+11 joint) axiom 6으로 pivot (sharper); Lemma 9 산출
2 Wall #2 codification ∫E(t)dt unconditional bound check Lemma 10 산출
3 Active program identification Lemmas 9, 10 ↔ Connes program mapping Lemma 3 update; partial unification 가설
4 Cycle 3 verification 4-paper Connes–Consani read Cycle 3 partial refutation; Wall #1 only, Wall #2 X
5 Path 1 active monitoring Selecta Math 2021 read 1년 direct progress documented
6 Lemma 9 falsifier test PNAS 2022 (Connes–Moscovici) 3 paper-direct gaps; Lemma 9 11/11로 강화
7 Externalization Preprint Section 1+2 draft (in progress)

주목할 패턴

  • Phase 1이 별도 attempt directory. 명시적 ideation 강제 (execution과 blend X).
  • Cycle ≠ attempt. Cycle은 multiple attempts span.
  • Codification cycles vs active monitoring cycles. Cycle 1, 2는 lemma 산출 (codification). Cycles 3, 4, 5는 기존 lemma에 paper-direct anchor 추가 (active monitoring). Cycle 6은 기존 lemma stress-test. 의도적 mix — anti-codification check (post-Critique #8)이 wall codification lemma mass-production 방지.

Transferable

Protocol은 RH-specific knowledge 가정 X. Scaffolding (intuition score, falsifier criterion, Specialist Δ paper-direct anchoring, anti-codification self-check)이 다른 “long-running LLM autonomous research on a hard problem”에 일반화.

Audit trail (Layer 1)

  • HARNESS.md §0
  • attempts/184_cycle1_* ~ attempts/190_cycle7_*
  • learnings/sustained_research_log.md
  • learnings/intuition_calibration.md

← Previous · English · Next →


AI-generated. Not a proof. RH-progress: 0/10. Contact: x2ever.han@gmail.com

This site uses Just the Docs, a documentation theme for Jekyll.