Vision Foresight Futures Wheel Delphi Rounds

Sub-skill: Delphi Wheel (5-round async + Wiki)

출처: Glenn, J. C. (2009). "Futures Wheel." Futures Research Methodology V3.0, Ch.6 §VI. Millennium Project. ISBN: 978-0-9818941-1-3. 상위 마스터: vision-foresight-futures-wheel 호출 권한: 마스터 orchestration 전용 (disable-model-invocation: true) Cross-skill: foresight-expert-pool · foresight-delphi · foresight-realtime-delphi · foresight-futures-polygon 결정론 모듈: delphi_utils.py (수치 계산 전량 위임 — LLM 자연어 추정 금지)

1. PDF 원전 정의

Glenn(2009) 직접 인용 (Frontiers section):

"Another future variation of the Futures Wheel could use a Delphi via the Internet. An international panel could assemble asynchronously to systematically construct a Futures Wheel:"

"Round 1: Ask an international panel to rate a list of events or trends for use with a Futures Wheel and or ask for additional suggestions;"

"Round 2: Feed back the panel's responses for further refinement, clarification, and ranking;"

"Round 3: Request respondents to list primary consequences of the trends or events of highest ranking;"

"Round 4: Display results as a Futures Wheel with just the primary ring of impacts; the size of the oval around each primary impact could represent the frequency with which the panel identified it; then ask for the secondary impacts;"

"Round 5: Display primary impacts as first ring, and secondary impacts as a second ring; again, the size of the ovals (or some other graphic device) could represent the frequency of responses."

"This approach could also use the ideas in the Futures Polygon for expressing consequences."

"These versions of the Futures Wheel could be assisted by collaborative software or groupware, which would collect and display the panel's views on the impacts in the graphic of a Futures Wheel."

"Futures Wheel Wikis could be created by letting geographically dispersed people add to and/or edit the consequences via the Internet."

2. AI Agent 6 종 구성

3. 8 Phase 처리 흐름

Phase 1 — Panel Recruitment

Async Panelists 5~30인을 foresight-expert-pool cross-skill로 캐스팅. 다양성 차원:

panel_diversity_dimensions:
  domain:
    - Technologist (예: Jerome Glenn 본인·Bishop·Coates·Voros)
    - Economist
    - Sociologist
    - Policy Maker
    - Industry Practitioner
    - Academia
    - Civic Society
  geography:
    - North America
    - Europe
    - East Asia (KR/JP/CN)
    - Global South
  perspective:
    - Optimist
    - Pessimist
    - Pragmatist
    - Critic
  age_cohort:
    - Boomer
    - Gen X
    - Millennial
    - Gen Z

최소 다양성 강제: 도메인 4종 이상, 지리 3종 이상, 관점 3종 이상.

[결정론 검증 — Phase 1 필수] 패널 확정 후 반드시 Python으로 다양성 점수 계산:

python3 delphi_utils.py diversity_score '<json_panelist_list>'

minimums_met: false 이면 추가 패널 리크루팅 없이 Round 1 진입 금지. score < 0.40 이면 마스터 오케스트레이터에 경고. → 상세: references/panel_diversity_matrix.md

Phase 2 — Round 1: List Rating + Suggestions

각 Async Panelist에게 anonymous 동시 제시:

## Round 1 (Anonymous)

다음은 Futures Wheel 분석 후보 trend·event list입니다.
각 항목에 대해 다음을 답해주세요:

1. **중요도 (1~5)**: Futures Wheel 분석 가치
2. **임팩트 (1~5)**: 잠재 사회 영향
3. **확실성 (1~5)**: 발생 가능성

추가로 제안하고 싶은 trend·event가 있으면 자유롭게 적어주세요.

### 후보 list
- [ ] {trend 1}
- [ ] {trend 2}
- [ ] {event 1}
- ...

Aggregator가 anonymous 응답 수집·집계.

[결정론 — Round 1 집계] 평균·표준편차·순위는 LLM이 추정하지 않고 Python으로 계산:

python3 delphi_utils.py aggregate_composite '<json_ratings_by_item>'

입력 형식: {"trend_name": {"importance": [4,5,3,...], "impact": [5,4,4,...], "certainty": [3,4,3,...]}} 출력: composite_mean, dim_means, std_dev_pooled, divergence_flag, top_candidate(bool) Top 후보 기준: composite_mean = avg(importance, impact, certainty) ≥ 4.0 (자동 ★ 표시)

Phase 3 — Round 2: Feedback + Refinement

Round 1 집계 결과를 panel에 redistribute:

## Round 2 (Anonymous)

Round 1 종합 결과를 공유합니다. 다음을 답해주세요:

1. 본인의 평가가 panel 평균과 다르면 *이유* 명시
2. 새 trend 제안 중 어떤 것이 가장 가치 있는지 ranking
3. trend·event 정의를 명료화하기 위한 refinement 제안

### Round 1 종합 (aggregate_composite_ratings() 출력 기반 자동 생성)
| Trend/Event | 평균 중요도 | 평균 임팩트 | 평균 확실성 | Top 후보 |
|-------------|-----------|-----------|-----------|---------|
| {trend_label_1} | 4.2 | 4.5 | 3.8 | ★ |
| {trend_label_2} | 3.8 | 3.5 | 4.1 |   |
| {trend_label_N} | ... | ... | ... | ... |

### 새로 제안된 trend (Round 1 panel 추가)
- {new trend A} (3 panelists 제안)
- {new trend B} (1 panelist 제안)

Phase 4 — Round 3: Primary Consequences

Round 2에서 highest-ranked trend·event 선정 후:

[결정론 — highest-ranked 선정] Round 2 재평가 후 순위 재계산:

python3 delphi_utils.py aggregate_composite '<updated_json>'

최고 composite_mean 항목을 Top Trend로 선정. 동점 시 std_dev_pooled 낮은 항목 우선 (합의도 기준).

## Round 3 (Anonymous)

Round 2 결과 panel이 가장 가치 있다고 동의한 trend는:
> **"{Top Trend}"** (composite score: {composite_mean}/5, panel consensus: {std_dev})

이 trend의 **primary consequence(직접 영향)**를 자유롭게 list해주세요.
- 시간 범위: T+1~5y
- 개수: 자유 (보통 3~10개)
- 형식: 짧은 phrase

Aggregator가 각 panelist의 primary list 수집.

Phase 5 — Round 4: Frequency-weighted Primary + Ask Secondary

Frequency Counter가 모든 primary impact 빈도 집계:

primary_frequency:
  "AGI로 화이트칼라 자동화": 28/30 panelists  # 93% — biggest oval
  "데이터센터 전력 수요 폭증": 25/30 panelists  # 83%
  "교육 시스템 재편": 22/30 panelists  # 73%
  "지정학 재편": 18/30 panelists  # 60%
  "AI governance 입법 가속": 15/30 panelists  # 50%
  # (further items...)

[결정론 — 빈도 계산 및 oval size 매핑] LLM이 추정하지 않고 반드시 Python 실행:

# 빈도 카운트 + oval size 자동 매핑
python3 delphi_utils.py count_frequencies '<json_list_of_lists>' <panel_size>

입력: 각 panelist의 primary impact 목록 (list of lists) 출력: {impact: {count, frequency, frequency_pct, oval_size}} 빈도 내림차순 → 상세 매핑 테이블: references/frequency_weighted_oval.md

빈도 → oval size 매핑 (PDF 명시, Glenn §VI Round 4):

이어서 panel에 redistribute:

## Round 4

다음은 panel이 식별한 primary impacts입니다 (oval 크기 = panel 동의 빈도).

[Frequency-weighted Wheel 시각화]

각 primary impact별로 **secondary impact(2차 영향)**를 list해주세요.
- 시간 범위: T+5~10y
- 개수: 각 primary마다 2~3개
- *각 primary의 secondary가 다른 primary 영향과 cross-link되는지* 의식하면서 작성

Phase 6 — Round 5: Frequency-weighted Primary + Secondary

Round 4 secondary 응답 수집 후 빈도 가중 (secondary 빈도도 동일하게 Python으로 계산):

[결정론 — Round 5 secondary 빈도]

python3 delphi_utils.py count_frequencies '<json_secondary_lists>' <panel_size>

primary별로 secondary를 별도로 집계. 예: P1 하위 secondary 응답만 모아 P1 secondary 빈도 산출.

## Round 5 (Final)

Panel이 합성한 Futures Wheel 최종본:

[2-ring Frequency-weighted Wheel 시각화]
- 1st ring: primary (Round 4 frequency size)
- 2nd ring: secondary (Round 5 frequency size)
- spoke: panel 동의 강도 (line thickness)
- (옵션) cross-link: cross-domain linkage 강조

본인의 최종 평가:
1. wheel의 plausibility (1~5)
2. 누락된 중요 영향 (있으면)
3. 추가 의견

Phase 7 — Wiki Collaboration

PDF 명시: "Futures Wheel Wikis could be created by letting geographically dispersed people add to and/or edit the consequences via the Internet."

Wiki Editor Agent가 다음 markdown wiki 구조 제공:

# Futures Wheel Wiki: {중앙 이슈}

[Latest version {CURRENT_DATETIME_UTC}]  ← python3 delphi_utils.py now_utc

## 1st Ring: Primary Impacts

### P1. AGI로 화이트칼라 자동화 [93% panel consensus]
- **Added by**: {panelist alias}
- **Last edited**: {timestamp}
- **Edit history**: [v1, v2 (refined wording), v3 (added time range)]
- **Discussion**:
  - {comment 1}
  - {comment 2}

### P2. ...

## 2nd Ring: Secondary Impacts

### S1a. (under P1) 신입 사무직 채용 시장 50% 축소 [78%]
- **Added by**: ...
- ...

## Cross-linkage (panel-suggested)
- S1a ←→ S3b (cross-domain link)
- S2a forms feedback loop with P2

## Open Issues
- [ ] P3 정의 panel 간 의견 차이 (Round 5)
- [ ] T+30y impact 까지 확장 필요?

기여자별 edit history 보존, 격리된 의견은 별도 섹션.

Phase 8 — Groupware Aggregation

PDF 명시: "These versions of the Futures Wheel could be assisted by collaborative software or groupware, which would collect and display the panel's views on the impacts in the graphic of a Futures Wheel."

최종 시각화에 다음 모두 표시:

final_visualization_layers:
  - oval_size: frequency (panel 합의도)
  - line_thickness: linkage strength
  - line_style: 
      solid: high consensus
      dashed: medium consensus
      dotted: low consensus / minority opinion
  - color:
      green: positive impact 다수 의견
      red: negative impact 다수 의견
      yellow: 양면·논쟁
  - icon:
      ⚠️: divergence (panel 큰 의견 차이)
      🔁: feedback loop 참여
      ⚡: critical issue (contradiction)

4. Frequency-weighted Visualization Templates

4.1 Mermaid (size 근사 표현)

flowchart TD
    C(("**중앙 이슈**"))

    C ===> P1["**P1** [93%]<br>AGI 화이트칼라 자동화"]
    C ===> P2["**P2** [83%]<br>전력 수요 폭증"]
    C ==> P3["P3 [73%]<br>교육 재편"]
    C --> P4["P4 [60%]<br>지정학 재편"]
    C -.-> P5["P5 [50%]<br>AI governance"]

    P1 ===> S1a["**S1a** [78%]<br>신입 채용 50% 축소"]
    P1 ==> S1b["S1b [65%]<br>인문·예술 가치 재상승"]
    P2 ==> S2a["S2a [70%]<br>SMR 산업 부상"]
    P2 --> S2b["S2b [55%]<br>에너지 지정학 재편"]
    P3 --> S3a["S3a [60%]<br>직업교육 패러다임 전환"]
    P4 -.-> S4a["S4a [35%]<br>AI 군비경쟁"]
    %% P5 secondary impacts to be added per actual panel data

    classDef xl fill:#1b5e20,color:#fff,stroke-width:5px
    classDef high fill:#2d6a4f,color:#fff,stroke-width:4px
    classDef medium fill:#457b9d,color:#fff,stroke-width:3px
    classDef low fill:#888,color:#fff,stroke-width:2px
    classDef negative fill:#d62828,color:#fff,stroke-width:3px
    classDef contested fill:#f4a261,color:#000,stroke-width:2px
    class P1 xl
    class P2,S1a high
    class P3,S1b,S2a medium
    class P4,P5,S3a low
    class S2b contested
    class S4a negative
    %% Color guide: xl/high/medium=positive, negative=red(harm), contested=amber(mixed)

4.2 ASCII Frequency-weighted

                    ┌────────────────┐
                    │   P1 [93%]     │  ← 가장 큰 oval
                    │  AGI 자동화    │
                    └───────┬────────┘
                            │
                ┌───────────┴────────────┐
        ┌───────┴────────┐        ┌──────┴───────┐
        │  S1a [78%]     │        │  S1b [65%]   │
        │  신입 50%↓     │        │  인문 재부상  │
        └────────────────┘        └──────────────┘
        
        ┌──────────────┐
        │  P2 [83%]    │
        │  전력 수요   │
        └──────┬───────┘
               │
        ┌──────┴──────┐
        │ S2a [70%]   │
        │ SMR 부상    │
        └─────────────┘
        
        ╭─ P3 [73%] ─╮
        │ 교육 재편  │
        ╰────────────╯
        
        · P5 [50%] ·  ← 작은 oval (낮은 consensus)
        AI governance

4.3 Panel Divergence Heatmap

| Impact | Panel Mean | Std Dev | Divergence Flag |
|--------|-----------|---------|----------------|
| P1 AGI 자동화 | 4.6/5 | 0.4 | ✓ consensus         |
| P2 전력 | 4.2 | 0.6 | ⚠️ moderate divergence |
| P3 교육 | 4.0 | 0.9 | ⚠️ moderate divergence |
| P4 지정학 | 3.5 | 1.4 | 🚨 high divergence    |
| P5 governance | 3.0 | 1.6 | 🚨 high divergence   |

임계값: std < 0.6 → ✓ consensus · 0.6 ≤ std < 1.2 → ⚠️ moderate · std ≥ 1.2 → 🚨 high (모든 분류는 delphi_utils.py::_classify_divergence() 함수로 계산 — LLM 추정 금지)

Divergence flag가 높은 항목은 minority opinion 으로 별도 보관.

5. PDF 인용 fragment

5.1 1차 출처 (Glenn 2009)

정식 서지: Glenn, Jerome C. (2009). "Futures Wheel." In Glenn, J. C., Gordon, T. J., & Florescu, E. (Eds.), Futures Research Methodology — Version 3.0. The Millennium Project, American Council for the United Nations University. Chapter 6, Section VI "Frontiers of the Method." ISBN: 978-0-9818941-1-3.

"Another future variation of the Futures Wheel could use a Delphi via the Internet. An international panel could assemble asynchronously to systematically construct a Futures Wheel." (§VI)

"Round 1: Ask an international panel to rate a list of events or trends for use with a Futures Wheel and or ask for additional suggestions." (§VI)

"Round 2: Feed back the panel's responses for further refinement, clarification, and ranking." (§VI)

"Round 3: Request respondents to list primary consequences of the trends or events of highest ranking." (§VI)

"Round 4: Display results as a Futures Wheel with just the primary ring of impacts; the size of the oval around each primary impact could represent the frequency with which the panel identified it; then ask for the secondary impacts." (§VI)

"Round 5: Display primary impacts as first ring, and secondary impacts as a second ring; again, the size of the ovals (or some other graphic device) could represent the frequency of responses." (§VI)

"This approach could also use the ideas in the Futures Polygon for expressing consequences." (§VI)

"These versions of the Futures Wheel could be assisted by collaborative software or groupware, which would collect and display the panel's views on the impacts in the graphic of a Futures Wheel." (§VI)

"Futures Wheel Wikis could be created by letting geographically dispersed people add to and/or edit the consequences via the Internet." (§VI)

5.2 Futures Polygon Cross-reference

Glenn (§VI) 명시: "This approach could also use the ideas in the Futures Polygon for expressing consequences."

• Round 4·5에서 식별된 primary/secondary impact는 선택적으로 Futures Polygon 포맷으로 다차원 표현 가능
• Futures Polygon: probability, desirability, time horizon, breadth 등 다차원 vertices
• Cross-skill: foresight-futures-polygon → foresight-futures-polygon-fw-import 경로로 impact 데이터 전달
• → 상세: references/glenn_5round_delphi.md § "Futures Polygon Cross-Reference"

5.3 지원 학술 출처 (결정론 환원 불가 항목 검증용)

• Dalkey, N. C., & Helmer, O. (1963). "An Experimental Application of the Delphi Method to the Use of Experts." Management Science, 9(3), 458–467. — Delphi 방법론 원전, 익명성·feedback·iteration 원칙
• Linstone, H. A., & Turoff, M. (Eds.) (1975). The Delphi Method: Techniques and Applications. Addison-Wesley. — 패널 다양성·minority opinion 보존 원칙
• Gordon, T. J. (1994). The Delphi Method. Futures Research Methodology, AC/UNU. — 분산 기반 divergence 해석 원칙

6. 마스터 입력 인터페이스

sub_skill: vision-foresight-futures-wheel-delphi-rounds
inputs:
  center_issue: "{issue_title}"   # 예: "AGI 상용화", "기후 위기 임계점"
  panel_size: 5~30
  panel_diversity_targets:
    domain_min: 4
    geography_min: 3
    perspective_min: 3
  expert_pool_request:
    cross_skill: foresight-expert-pool
    mode: "REFRESH" | "READ" | "USER-EXPLICIT"
  rounds_to_run: 1~5   # 부분 round 실행 가능
  enable_wiki: true
  enable_groupware_aggregation: true
  frequency_thresholds: { XL: 0.9, L: 0.7, M: 0.5, S: 0.3, XS: 0.0 }
  # XS: <30% (lowest bucket, implicit fallback). Validate with delphi_utils.py::validate_frequency_thresholds()
  panel_size_constraint: { min: 5, max: 30, type: int }
  # Enforced by delphi_utils.py::validate_panel_size() before Round 1
outputs:
  - panel_recruitment_report       # Phase 1 다양성 점수 + panelist 목록
  - round_by_round_responses       # 각 Round의 anonymous aggregate 데이터
  - frequency_weighted_wheel       # Phase 5·6 Mermaid + ASCII 시각화
  - wiki_markdown                  # Phase 7 Wiki 전체 markdown
  - groupware_final_visualization  # Phase 8 레이어드 최종 시각화
  - panel_divergence_heatmap       # 4.3 divergence heatmap
  - minority_opinions_archive      # 🚨 high divergence 항목 전체
  - pdf_citations                  # Section 5 Glenn 인용 목록

7. 호출 후 마스터로 반환

sub_skill_output:
  status: completed | partial | error
  panel_recruited: N               # 실제 캐스팅된 panelist 수
  panel_diversity_score: 0.0~1.0   # python3 delphi_utils.py diversity_score 결과
  panel_diversity_minimums_met: true | false
  rounds_completed: 1~5
  primary_consensus_top:           # Round 4 기준 상위 primary impacts
    - item: "..."
      frequency: 0.93
      oval_size: "XL"
  secondary_consensus_top:         # Round 5 기준 상위 secondary impacts
    - item: "..."
      under: "P1"
      frequency: 0.78
      oval_size: "L"
  divergence_flags:                # 🚨 high divergence 항목 전체
    - item: "..."
      mean: 3.2
      std_dev: 1.4
      flag: "🚨 high divergence"
      archive_id: "MO-{n}"
      escalate_to: "foresight-scenario-forecast"
  wiki_url_or_md_path: "..."
  visualizations:
    frequency_weighted: "mermaid block or ascii"
    divergence_heatmap: "markdown table"
    groupware_final: "layered visualization block"
  panel_recruitment_report: "..."
  round_by_round_responses: [...]
  minority_opinions_archive: [...]
  pdf_citations:
    - "Glenn, J. C. (2009). Futures Wheel. §VI. Millennium Project."
    - "Dalkey, N. C., & Helmer, O. (1963). Management Science, 9(3), 458–467."
    - "Linstone, H. A., & Turoff, M. (Eds.) (1975). The Delphi Method. Addison-Wesley."

마스터는 divergence_flags가 높은 항목을 foresight-scenario-forecast로 escalate (시나리오 분기점).

8. Cross-skill 호환

본 sub-skill은 다음 cross-skill과 호환:

9. references/

10. 결정론 환원 모듈 (delphi_utils.py)

LLM이 자연어로 추론하면 할루시네이션 위험이 있는 모든 수치 계산은 delphi_utils.py로 위임한다.

결정론 환원 불가 항목 — 학술 출처로 1:1 대조 검증:

11. 오류 및 예외 처리