Designing and Evaluating a Learning-Centric Scrum-Embedded Agile Requirements Engineering Framework : Uncertainty-Aware Routing and Inspectable Decision Traces
- 주제(키워드) agile requirements engineering , uncertainty-aware routing , value of uncertainty and learning （VOUL） , inspectable decision traces , Scrum-embedded requirements engineering
- 발행기관 아주대학교 정보통신전문대학원
- 지도교수 Seok-Won Lee
- 발행년도 2026
- 학위수여년월 2026. 8
- 학위명 박사
- 학과 및 전공 정보통신전문대학원 정보통신공학전공
- 실제URI http://www.dcollection.net/handler/ajou/000000036610
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Agile RE 팀은 두 가지 요구공학 문제를 안고 있다. 불확실성이 높은 백로그 항목이 증거 상태나 경로 결정 근거가 명시되기 전에 delivery 방향으로 진행되면서 해소되지 않은 가설이 구현 작업에 내재되고, 설령 Sprint Review에서 올바른 결정이 내려지더라도 그 근거가 스프린트 경계를 넘어 나중 독자가 재구성할 수 있는 형태로 보존되는 경우는 드물다. 기존 Agile RE 연구는 문제 프레이밍, 가설 기반 개발, 추적성을 각각 다루었으나, 세 가지를 Scrum 안에서 거버넌스 부담 없이 연결하는 메커니즘은 부재했다. 본 논문은 Value of Uncertainty and Learning（VOUL） 원리에 기반한 Scrum 내재형 경로-검토 프레임워크를 제안한다. VOUL은 의사결정 분석의 정보 가치（VOI） 이론을 저자가 Agile RE 실무에 맞도록 운용화한 원리로, VOI가 결정을 확정하기 전 추가 정보 획득의 기대 이득을 묻는다면, VOUL은 그 논리를 스프린트 단위 운용 규칙으로 전환한다. 백로그 항목의 불확실성은 빠른 구현으로 억제할 위험이 아니라 목표 지향적 학습의 신호로 다룬다. 프레임워크는 이 원리를 기존 Scrum 이벤트 안에 세 가지 경량 결정 시점으로 구현한다. Backlog Refinement에서는 문제 중요도와 증거 강도를 점수화해 권고 경로（SuggestedRoute）를 산출하고, 팀은 실제로 적용한 Track을 별도로 기록한다. Sprint Planning에서는 결과를 알기 전에 항목의 build-first 의도를 기록한다. Sprint Review에서는 GO, PIVOT, KILL 처분을 증거 링크 및 결과 요약과 함께 기록하며, PIVOT은 실패가 아닌 수정된 가설로 학습을 보존하는 결정으로 정의된다. 새로운 Scrum 이벤트는 추가되지 않는다. 프레임워크는 세 개 산업 Scrum 팀, 6회 2주 스프린트, 319개 고유 백로그 항목을 대상으로 평가됐다. 최종적으로 non-GO로 종결된 79개 항목 중 70개가 스프린트 종료 이전에 PIVOT 또는 KILL에 도달했다（최신 항목 관점 88.6%, 항목별 스프린트 패널 87.6%）. 이것이 핵심 연구 질문의 결과다. 불확실성이 높은 항목이 delivery commitment 이전에 스프린트 내에서 명시적인 PIVOT 또는 KILL 처분으로 재방향됐다. Sprint Planning 시점에 팀원들이 delivery 가능성이 있다고 판단한 135개 항목 중 47개（34.8%）가 명시적 증거 검토 후 재방향됐다는 사실은 이 격차를 더 구체적으로 보여 준다. 이월된 53개 carry-over 스냅샷 전체가 사후 검사에 필요한 최소 연결 필드 집합을 보존했으며, Sprints 7~8 포함 시 68/68로 유지됐다. 이러한 결과들은 경로 결정을 명시적으로 기록하고 스프린트 경계를 넘어 검사할 수 있음을 보여 주며, 이론적으로는 Agile RE를 불확실성 아래의 요구사항 단위 inspect-and-adapt로 재프레이밍하는 동시에 routing과 ranking을 별개의 결정으로 분리하는 데 기여한다. 연구의 기여는 의도적으로 한정된다. 결과는 경로 결정 메커니즘의 분석적 일반화를 지지하며, 통계적 일반화나 인과적 증명을 주장하지 않는다. 향후 연구로는 저자가 직접 개입하지 않는 독립적 반복 연구, 기존 백로그 도구와의 통합, 종단 연구가 필요하다. Sprint Review에서 생성된 구조화된 근거 링크는 AI 지원 검색 및 요약의 입력 자료로 활용하는 방향도 탐색할 수 있다. 이론적 기여의 핵심은 ranking과 routing의 구분에 있다. ranking은 항목의 우선순위를 정하고, routing은 해당 항목에 지금 어떤 작업이 필요한지를 결정한다. 이 논문은 Agile RE를 불확실성하에서의 요구사항 수준 inspect-and-adapt로 재프레이밍하며, 두 결정을 명시적으로 분리하고 그 분리를 단일 검사 가능한 레코드로 운용화한다.
more초록/요약
Agile software development teams carry two persistent requirement-level problems into every sprint. Uncertainty-rich backlog items move toward delivery commitment before their evidence state or routing rationale has been made explicit, converting unresolved assumptions into implementation work. And even when a sound routing decision is made at Sprint Review, the reasoning behind it rarely survives the sprint boundary in a form that later readers can reconstruct. Prior Agile RE research has addressed problem framing, hypothesis-driven development, and traceability in isolation, but has not produced a mechanism that connects all three inside routine Scrum practice without adding governance overhead. This dissertation proposes a Scrum-embedded route-and-review framework guided by the Value of Uncertainty and Learning （VOUL） principle. VOUL is the author's operational adaptation of value-of-information （VOI） reasoning from decision analysis to Agile RE practice. Where formal VOI asks what the expected gain from additional information is before committing to a decision, VOUL translates that logic into a sprint-level rule: uncertainty in a backlog item is a signal for targeted learning, not a risk to suppress through faster delivery. The framework acts on this principle by inserting three lightweight decision moments into existing Scrum events. At Backlog Refinement, each item receives a problem-importance score and an evidence-strength score that generate an advisory route recommendation; the team separately records the Track it enacts. At Sprint Planning, the team records whether the item carries build-first momentum before any review outcome is known. At Sprint Review, an explicit GO, PIVOT, or KILL disposition is recorded alongside evidence links and an outcome summary. PIVOT is not a failure label; it preserves the item as a learning object with a revised hypothesis. No new Scrum ceremony is introduced, and no existing event is restructured. The framework was evaluated with three industrial Scrum teams across six two-week sprints, covering 319 unique backlog items. Among 79 items that ultimately ended as non-GO decisions, 70 reached PIVOT or KILL before sprint closure, an earlier-redirection rate of 88.6% in the latest- item view and 87.6% in the item-by-sprint panel. This is the core finding for the primary research question: uncertainty-rich items were surfaced and redirected within the sprint, before delivery commitment hardened, rather than being carried silently into implementation. A secondary analysis sharpens the picture further: among 135 items that practitioners themselves judged as plausible pull-in candidates at Sprint Planning, 47 （34.8%） were ultimately redirected or stopped once explicit evidence review was applied. The gap between that implicit readiness judgment and the criterion-based outcome is precisely what the protocol makes visible. For trace continuity, all 53 carry-over snapshots across sprint boundaries retained the minimum linked field set required for field-complete reconstructability of learning decision traces, a result that held at 68 of 68 snapshots when Sprints 7 and 8 were included. Together, these results show that routing decisions can be made explicit, recorded in a lightweight linked form, and inspected across sprint boundaries within ordinary Scrum work, without adding a new recurring Scrum ceremony. Ranking orders items; routing determines what kind of work should happen next. The theoretical contribution reframes Agile RE as requirement-level inspect-and-adapt under uncertainty, treating routing and ranking as distinct decisions and operationalizing that separation through a minimum field set that connects framing, evidence state, and review disposition in a single inspectable record. The contribution is intentionally bounded. The results support analytical generalization of the routing mechanism, not statistical generalization to all Agile organizations, and no causal proof is offered that fewer failed deliveries result from adopting the protocol. Future priorities include independent replication without author facilitation, integration with existing backlog tools, and longitudinal adoption studies to determine whether routing behaviors persist once the embedded- researcher scaffold is removed. The structured rationale links produced during Sprint Review also open a near-term direction for LLM-assisted retrieval and documentation generation, where preserved decision traces serve as structured inputs to summarization.
more목차
I. Introduction 1
A. Background and Motivation 1
B. Problem Statement 2
C. Research Goal and Scope 4
D. Research Questions 5
E. Overview of the Proposed Framework 7
F. Empirical Setting and Evidence Structure 9
G. Contributions 10
H. Dissertation Organization 11
II. Literature Review 13
A. Persistent Agile RE Challenges under Uncertainty 13
B. Problem Framing and Requirement Representation 15
C. Hypothesis Engineering and Experiment-Driven Product Development 18
D. Dual-Track and Continuous Experimentation 21
E. Value and Uncertainty in Software Decisions 24
F. Continuous Engineering and Scrum-Embedded Learning 28
G. Traceability, Rationale, and Inspectability 31
H. Literature-to-Framework Synthesis and Gap-to-RQ Preparation 36
III. Research Design 42
A. Research Problem and Goal 42
B. Research Questions and Proposition Structure 43
C. Research Design Overview 47
D. Units of Analysis and Evidence Roles 49
E. Data Sources and Analytic Views 53
F. Claim Discipline and Validity Boundaries 60
G. Summary 66
IV. Integrated Framework and VOUL-Guided Protocol 68
A. Design Objective and Intervention Boundary 68
B. Overall Framework Architecture 70
C. Why VOUL Is Used Here 71
D. Upstream Framing Mechanism 73
E. Core Route-and-Review Protocol 74
1. Delivery-Track Handling Boundary 78
F. Field Economy and Minimal-Burden Design 78
G. Transition Rule and Calibration 80
H. Scrum Embedding Logic 82
I. Decision Traces, Inspectability, and Reuse 84
J. Summary 86
V. Evaluation Methodology and Data Operationalization 88
A. Research Goal and Evaluation Scope 88
B. Research Design 89
C. Setting and Evaluation Windows 90
D. Units and Evidence Mapping 92
E. Dataset and Analytic Views 95
F. Reviewer-Reconstruction Component 96
G. Data Handling and Provenance 98
H. Validity Controls 102
1. Anchoring Workshop Calibration 104
2. Researcher Positionality 104
3. Hawthorne and Compliance Effects 105
I. Ethics and Disclosure Boundary 105
J. Reporting Logic 106
K. Summary 107
VI. Core Empirical Evaluation （Sprints 1–6）: Early Redirection and Inspectable Traces 109
A. Chapter Purpose and Claim Hierarchy 109
B. Dataset and Analytic Views 110
C. Core Result 1: Earlier Redirection （RQ3） 113
1. Primary Prevalence Statement 113
2. Score Profile Separability 113
3. Timing Corroboration: Panel View 114
4. Planning-Time Anchor: The BaselinePullIn Slice 116
5. Team-Level Robustness and Heterogeneity 117
6. Record-Level Illustration 119
7. What RQ3 Establishes 120
D. Core Result 2: Trace Continuity （RQ4） 120
1. The Carry-Over Observation Window 120
2. Traceability Versus Inspectability 121
3. Two Illustrative Traces 122
4. What RQ4 Establishes 122
E. Supporting Result 1: NFR Visibility （RQ1） 122
F. Supporting Result 2: Hypothesis Visibility （RQ2） 123
G. Supporting Result 3: Downstream Availability （RQ5） 125
H. Validity Boundaries 130
I. Summary 132
VII. Extended Observation: Stability, Maturation, and Practical Realism （Sprints 7–8） 134
A. Purpose of Extended Observation 134
B. Sprints 7–8 Dataset Overview 134
C. Stability of Earlier Redirection 138
1. Panel and Latest-Window Stability 138
2. BaselinePullIn-Conditioned Slice 139
3. Sprint-by-Sprint Breakdown 140
D. PIVOT Maturation Evidence 141
1. S6–S7 Bridge Transitions 141
2. Full S6–S7 Bridge Items 141
3. Carry-Over Trace Continuity: Sprints 7–8 142
E. Extended Window Vignettes 143
F. Team-Level Heterogeneity 144
G. Field Coverage Sustainability 145
H. Extended Window Synthesis 147
I. Summary 148
VIII. Operationalization and Practitioner Runbook 150
A. Operationalization Principles 150
B. Roles and Responsibilities 151
C. Event-Level Runbook 152
D. Minimum Field Set and Derived Outputs 153
E. Decision Rules: GO, PIVOT, and KILL 154
F. Evidence and OutcomeSummary Rules 156
G. Adoption Friction and Resistance Handling 157
H. Data Extraction and Analysis Handoff 158
I. Runbook Checklist 159
J. Worked Decision Examples 160
K. OutcomeSummary Quality Guide 161
L. Deployment Modes 161
M. Failure Modes and Countermeasures 162
N. Practitioner Burden and Field Economy 163
O. Summary 164
IX. Discussion 166
A. Interpretive Frame and Claim Hierarchy 166
B. Core Mechanism: Redirection and Trace Continuity 167
C. Supporting Findings Overview 170
D. Theoretical Contributions 174
E. Methodological Contributions 178
F. Practical Implications 180
1. Agile RE as Learning Discipline 180
2. Implications for Scrum Teams 181
3. Implications for Product Owners 181
4. Implications for Requirements Analysts 181
5. Implications for Developers and Sprint Execution 181
6. Implications for Evidence-Based Practice 182
7. Implications for Toolchain Design 182
8. Implications for NFR Handling 182
9. Implications for AI-Assisted Agile RE 182
10. Adoption Guidance for Practitioners 182
11. Boundary Conditions and Practical Cautions 183
12. Summary of Practical Implications 183
G. Limitations and Boundary Conditions 183
1. Why Boundary Conditions Matter 183
2. Causal Interpretation and BaselinePullIn 184
3. Measurement, Calibration, and Thresholds 184
4. Trace Inspectability and Denominator Discipline 185
5. Supporting Findings and External Validity 186
6. Tooling, AI, Long-Term Outcomes, and Culture 186
H. Summary 187
X. Conclusion and Future Work 190
A. Problem and Claim 190
B. Findings 190
C. Implications 192
D. Practice 193
E. Boundaries 194
F. Future Work 195
G. Closing 196
References 198
Appendix A. Practitioner Runbook （Field Reference Version） 203
A.1 Purpose and scope 203
A.2 Minimum field set 203
A.3 Pre-sprint calibration 204
A.4 Sprint-by-sprint checklist 204
Backlog refinement 204
Sprint Planning 204
During the sprint 205
Sprint Review 205
Retrospective 205
A.5 GO/PIVOT/KILL decision guide 205
A.6 Scoring quick reference 206
A.7 OutcomeSummary writing guide 206
A.8 Common adoption failure modes 207
A.9 Relationship to the dissertation body 207
A.10 Adoption scenarios 207
Appendix B. Evaluation Instruments 209
B.1 Purpose and scope 209
B.2 BelScore calibration sheet 209
B.3 PlScore calibration sheet 210
B.4 ImpactScore calibration sheet 211
B.5 ExpCostScore and ReachScore calibration sheets 211
B.6 Sprint Planning anchor question list 212
B.7 Sprint Review recording form 213
B.8 RQ5 reviewer reconstruction form used in the bounded complement 214
B.9 Instrument use summary 215
B.10 Relationship to the dissertation body 215
Appendix C. Anonymized Dataset Summary 216
C.1 Purpose and disclosure boundary 216
C.2 Sprints 1–6 — Primary evaluation window 216
C.3 Sprints 7–8 — Extended observation window 218
C.4 Sprints 1–8 — Synthesis view 219
C.5 Score distribution summary 220
C.6 Relationship to the dissertation body 221
Apendix D. Field Schema and Derivation Logic 222
D.1 Purpose and scope 222
D.2 Field layers 222
D.3 Observed field list 223
D.4 Derived indicators 223
GapScore 223
TransitionReady 224
EarlyDecision 224
Learning efficiency proxy 224
Discovery Priority Score 224
D.5 Route classification rules 225
D.6 Denominator definition rules 225
Latest-item view 225
Item-by-sprint panel view 225
Carry-over view 226
Extension and synthesis views 226
D.7 BaselinePullIn definition and boundary 226
D.8 Public and internal schema distinction 226
D.9 Provenance path 227
D.10 Relationship to the dissertation body 227
Appendix E. Ethics and Disclosure Note 228
E.1 Purpose 228
E.2 Participant consent and anonymization 228
E.3 Organizational approval boundary 228
E.4 Data retention and access 229
E.5 AI disclosure 229
Appendix F. Glossary of Key Terms 230
F.1 Purpose 230
F.2 Glossary entries 230
F.3 Reader note 233
국문초록 234
