Clarify Graph-FL conclusion outcomes

Author: 2-mean

README.md

@@ -25,6 +25,16 @@ Graph-FL gain
 + optimizer effect
 ```
 
+## Conclusion Matrix
+
+이 framework의 주장은 “항상 graph가 필요하다”가 아니다. 실험 결론이 어느 쪽으로 나오든 다음 결정을 할 수 있게 만드는 것이 핵심이다.
+
+| 최종 결론 | Evidence pattern | 우리가 할 수 있는 결정 | 남는 framework 가치 |
+|---|---|---|---|
+| graph가 필수적이다 | real graph가 matched controls보다 일관되게 높고, real-control gap이 alignment/`LOO`/graph stats와 함께 움직임 | `graph_source`, `graph_mode`, `aggregation_target`을 method design 축으로 채택하고 Graph-FL mechanism을 강화 | graph gain을 relation-specific effect로 설명하고 prior work mechanism과 연결 |
+| graph가 필수적이지 않다 | graph-free correction, clustering-only, uniform/identity control이 real graph 효과를 대부분 설명 | graph construction을 줄이고 dominance/norm correction 또는 simpler control-based baseline으로 간소화 | Graph-FL gain이 graph 때문이 아니라는 결론을 artifact로 증명 |
+| 결론이 setting-dependent다 | seed, alpha, client count, graph_source, graph_mode에 따라 real-control gap이 달라짐 | 조건별로 Graph-FL 적용 구간을 제한하고 Evidence pattern 기준으로 method selection | 애매한 결과를 실패가 아니라 boundary와 decision rule로 정리 |
+
 ## Repository Contents
 
 | Area | Role | Main Paths |

docs/README.md

@@ -7,7 +7,7 @@
 | Document | 답하는 질문 | 핵심 근거 |
 |---|---|---|
 | `framework.md` | Graph-FL gain을 어떤 component와 metric으로 분해하는가 | lifecycle, `GraphFLDesign`, metric formula, artifact fields |
-| `evidence.md` | framework가 왜 정당한가 | experiment case matrix, 18 graph modes parity, 8,640 design-space checks, 4 / 4 extension checks |
+| `evidence.md` | framework가 왜 정당한가 | 18 graph modes parity, 8,640 design-space checks, 4 / 4 extension checks |
 | `research.md` | prior work와 어떤 위치 관계를 갖는가 | FedAMP, SFL, pFedGraph, FedAGA mapping |
 | `repository.md` | code/config/script/test/docs가 어디에 있는가 | repository tree, package layout, change routing |
 | `maintenance.md` | public surface와 compatibility 기준은 무엇인가 | canonical surface, alias table, golden baseline policy |

docs/evidence.md

@@ -40,34 +40,6 @@
 | Extensibility | custom source, builder, preset, target이 trace와 artifact까지 도달한다 | 4 / 4 extension rows pass | `extension_contract_summary.csv` |
 | Measurement integrity | real diagnostic과 expected-zero control을 구분한다 | real/random/uniform measured nonzero, identity expected-zero control | `real_diagnostic_consistency.csv` |
 
-## Experiment Case Matrix
-
-이 표는 실험 결과를 경우의 수별로 읽기 위한 index다. Historical case는 초기 실험 관찰을 보존하고, validation case는 현재 framework가 같은 claim을 artifact contract로 검증하는 방식이다.
-
-| Case Family | Cases | Main Variables | Result Signal | Current Artifact |
-|---|---|---|---|---|
-| Phase 1 diagnostics | Fashion-MNIST Dirichlet, 5 clients, seeds 42/43/44 | update norm, `DI`, `N_eff`, alignment, `LOO` | client contribution imbalance, seed variation, interaction pathology signal | `round_metrics.csv`, `client_metrics.csv` |
-| Phase 2 graph informativeness | update, random, shuffled, uniform, identity variants | classifier-head update source, control graph family | weak control boundary, update graph sensitivity | `counterfactual_metrics.csv`, `metric_validity_summary.csv` |
-| Phase 2 signed conflict | `signed_conflict_knn`, dense graph, real/update controls | graph_mode, graph density, control family | update graph effect limited, source/mode attribution needed | `graph_parity_summary.csv`, `composability_matrix.csv` |
-| Phase 2 client-count/alpha stress | n20 alpha 0.1, n20 alpha 0.3 | client count, Dirichlet alpha, `signed_conflict_knn` | graph-control separation weak and alpha-sensitive | `design_space_matrix.csv`, stress configs |
-| Phase 2 pFedSim-like | pFedSim-like graph, real/update controls | proxy graph quality, relation mechanism | control gap limited, paper-mechanism alignment needed | `external_mechanism_alignment.csv` |
-| Phase 2 source sanity | `update`, `ema_update`, `classifier_head_update` | graph_source, update/random/identity variants | source별 metric 차이와 artifact 생성 확인 | source contract, artifact contract |
-| Phase 2.5 smoothing | graph smoothing without relation signal | smoothing-only path, random/uniform controls | smoothing-only effect가 metric을 움직임 | `counterfactual_metrics.csv`, `real_diagnostic_consistency.csv` |
-| Phase 3 dominance-aware | dominance-aware baseline | accuracy, loss, update norm, `DI`, `N_eff` | dominance correction signal 확인 | graph-free correction diagnostics |
-| Framework parity | 18 graph modes | lifecycle assembly vs reference builder | max abs diff `2.21e-12`, edge F1 `1.0` | `graph_parity_summary.csv` |
-| Framework design-space | 16 sources x 18 modes x 5 targets x 6 correction profiles | source, mode, target, correction profile | 8,640 / 8,640 calculation checks pass | `design_space_matrix.csv`, `design_space_summary.csv` |
-| Framework extension | custom source, builder, preset, target | extension point, trace, metadata, artifact | 4 / 4 extension rows pass | `extension_contract_summary.csv` |
-
-Case 축은 다음처럼 해석한다.
-
-| Axis | Values Captured |
-|---|---|
-| graph source | update, EMA update, classifier-head update, weight, built-in source family |
-| graph mode | cosine/kNN, magnitude/RBF, learned smoothness, pFedGraph-QP, signed conflict variants |
-| aggregation target | `graph_filtered_update`, `graph_filtered_ema_update`, `graph_filtered_weight`, target variants |
-| control/correction | random, shuffled, uniform, identity, clustering-only, graph-free correction |
-| diagnostic metric | accuracy, loss, update norm, alignment, `DI`, `N_eff`, `LOO`, graph stats |
-
 ## Pass Criteria
 
 | Validation Unit | Pass Condition | Failure Signal |