C.E.L_Slide_test2/tests/matching/MATRIX_PHASE25.md

Phase 25 — Template-fit-v1 (최종 운영 체계)

공식: 0.25 anchor + 0.20 card + 0.20 rel + 0.15 slot + 0.20 content - penalties 결과: 4/4 정답

✓ 검증 기준 통과: 4/4, 오답 use_as_is 없음, 1-2위 격차 모두 ≥ 0.05

Phase 24 대비 업그레이드:

• DB: legacy 32 frames (family/semantic_role) → templates_v1 32개 (slot + anchor_sets + fit_notes + adaptation_allowed)
• 구조 축: 단일 → 3개 분해 (cardinality / relation / slot_coverage)
• 감점 축 2개: not_suits (의미 mismatch), adaptation_cost (재구성 비용)
• 조건부 cap: bim_dx, safety_quality_productivity 등 짧은 generic 세트 편향 방어
• 라우팅: use_as_is / light_edit / restructure / reject 4단계

1. TARGET별 결과

MDX	정답	1위 conf (라우팅)	2위 conf	margin
MDX01-2-details	18 (bim_dx_comparison_table)	18 (0.922 use_as_is) ✓	29 (0.847)	0.076
MDX02-2.2-table	14 (three_persona_benefits)	14 (0.928 use_as_is) ✓	20 (0.748)	0.180
MDX03-1	13 (three_parallel_requirements)	13 (0.925 use_as_is) ✓	20 (0.846)	0.079
MDX03-2	29 (process_product_two_way)	29 (0.919 use_as_is) ✓	18 (0.821)	0.098

2. Top-5 매트릭스 (32 템플릿 중)

콘텐츠	1위	2위	3위	4위	5위
MDX01-2-details 정답 18	🎯 18 (0.922) bim_dx_comparison_tabl use_as_is	29 (0.847) process_product_two_wa reject	24 (0.735) engn_sw_three_types restructure	23 (0.655) app_sw_package_vs_solu reject	31 (0.564) industry_characteristi reject
MDX02-2.2-table 정답 14	🎯 14 (0.928) three_persona_benefits use_as_is	20 (0.748) dx_sw_necessity_three_ restructure	21 (0.743) solution_engn_split_di restructure	13 (0.739) three_parallel_require reject	28 (0.722) sw_reality_three_empha reject
MDX03-1 정답 13	🎯 13 (0.925) three_parallel_require use_as_is	20 (0.846) dx_sw_necessity_three_ light_edit	28 (0.727) sw_reality_three_empha reject	21 (0.677) solution_engn_split_di reject	14 (0.659) three_persona_benefits reject
MDX03-2 정답 29	🎯 29 (0.919) process_product_two_wa use_as_is	18 (0.821) bim_dx_comparison_tabl reject	23 (0.645) app_sw_package_vs_solu reject	24 (0.641) engn_sw_three_types reject	27 (0.638) bim_adoption_central_s restructure

3. 유닛별 세부 (Top-5 breakdown)

MDX01-2-details — 정답 18 (bim_dx_comparison_table)

MDX 분석 (detect_mdx.py — LLM 0회)

• title: 건설산업 DX의 올바른 이해 / 2. 용어간 상호관계
• summary: 건설산업 DX의 올바른 이해 / 2. 용어간 상호관계. DX / BIM
• item_count: {'detected': 2, 'source': 'table_columns'}
• relation_type: {'value': 'compare', 'confidence': 'high'}
• detected_terms: 112개

🎯 1위 bim_dx_comparison_table (Frame 18) — use_as_is

• anchor: 1.00 — bim_dx: ['BIM', 'DX'] (cap0.70 면제 (방증 0.71≥0.5))
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.61
• base: 0.922
• confidence: 0.922 → use_as_is

2위 process_product_two_way (Frame 29) — reject

• anchor: 0.75 — dimension_2d_3d: ['2D', '3D', '3D모델']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.55
• base: 0.847
• confidence: 0.847 → reject

3위 engn_sw_three_types (Frame 24) — restructure

• anchor: 1.00 — sw_three_types: ['상용', '전용']
• cardinality: 0.50
• relation: 1.00
• slot: 0.50
• content: 0.55
• base: 0.735
• adaptation: -0.00 ['forbidden:n<min']
• confidence: 0.735 → restructure

4위 app_sw_package_vs_solution (Frame 23) — reject

• anchor: 0.00 — None: []
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.53
• base: 0.655
• confidence: 0.655 → reject

5위 industry_characteristics_three_col (Frame 31) — reject

• anchor: 0.33 — three_industries: ['토목']
• cardinality: 0.50
• relation: 1.00
• slot: 0.50
• content: 0.53
• base: 0.564
• adaptation: -0.00 ['forbidden:n<min']
• confidence: 0.564 → reject

MDX02-2.2-table — 정답 14 (three_persona_benefits)

MDX 분석 (detect_mdx.py — LLM 0회)

• title: DX의 시행 목표 및 기대효과 / 2. DX 기반 Process 혁신에 따른 주체별 기대효과 / 2.2 DX 시행 주체별 기대효과
• summary: DX의 시행 목표 및 기대효과 / 2. DX 기반 Process 혁신에 따른 주체별 기대효과 / 2.2 DX 시행 주체별 기대효과. 발주자 / 시공자 / 설계자
• item_count: {'detected': 3, 'source': 'table_columns'}
• relation_type: {'value': 'parallel', 'confidence': 'high'}
• detected_terms: 126개

🎯 1위 three_persona_benefits (Frame 14) — use_as_is

• anchor: 1.00 — construction_personas: ['발주자', '설계자', '시공자']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.64
• base: 0.928
• confidence: 0.928 → use_as_is

2위 dx_sw_necessity_three_perspectives (Frame 20) — restructure

• anchor: 0.33 — three_design_perspectives: ['디지털']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.57
• base: 0.748
• confidence: 0.748 → restructure

3위 solution_engn_split_diagram (Frame 21) — restructure

• anchor: 0.50 — solution_concepts: ['시공', '협업']
• cardinality: 0.80
• relation: 1.00
• slot: 1.00
• content: 0.54
• base: 0.743
• confidence: 0.743 → restructure

4위 three_parallel_requirements (Frame 13) — reject

• anchor: 0.33 — three_pillars: ['기술']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.53
• base: 0.739
• confidence: 0.739 → reject

5위 sw_reality_three_emphasis (Frame 28) — reject

• anchor: 0.33 — domain_civil: ['실무']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.45
• base: 0.722
• confidence: 0.722 → reject

MDX03-1 — 정답 13 (three_parallel_requirements)

MDX 분석 (detect_mdx.py — LLM 0회)

• title: DX 실행 체계 구축 방안 / 1. DX 시행을 위한 필수 요건
• summary: DX 실행 체계 구축 방안 / 1. DX 시행을 위한 필수 요건. 기술 / 사람 / 자연
• item_count: {'detected': 3, 'source': 'bullets'}
• relation_type: {'value': 'parallel', 'confidence': 'high'}
• detected_terms: 90개

🎯 1위 three_parallel_requirements (Frame 13) — use_as_is

• anchor: 1.00 — three_pillars: ['기술', '사람', '자연']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.63
• base: 0.925
• confidence: 0.925 → use_as_is

2위 dx_sw_necessity_three_perspectives (Frame 20) — light_edit

• anchor: 0.75 — dx_sw_necessity: ['DX', 'S/W', '필수']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.54
• base: 0.846
• confidence: 0.846 → light_edit

3위 sw_reality_three_emphasis (Frame 28) — reject

• anchor: 0.33 — domain_civil: ['토목']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.47
• base: 0.727
• confidence: 0.727 → reject

4위 solution_engn_split_diagram (Frame 21) — reject

• anchor: 0.25 — solution_engn_sw: ['S/W']
• cardinality: 0.80
• relation: 1.00
• slot: 1.00
• content: 0.52
• base: 0.677
• confidence: 0.677 → reject

5위 three_persona_benefits (Frame 14) — reject

• anchor: 0.00 — None: []
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.55
• base: 0.659
• confidence: 0.659 → reject

MDX03-2 — 정답 29 (process_product_two_way)

MDX 분석 (detect_mdx.py — LLM 0회)

• title: DX 실행 체계 구축 방안 / 2. Process의 혁신과 Product의 변화
• summary: DX 실행 체계 구축 방안 / 2. Process의 혁신과 Product의 변화. 과정 / 결과
• item_count: {'detected': 2, 'source': 'subsections'}
• relation_type: {'value': 'compare', 'confidence': 'high'}
• detected_terms: 117개

🎯 1위 process_product_two_way (Frame 29) — use_as_is

• anchor: 1.00 — dimension_2d_3d: ['2D', '2D도면', '3D', '3D모델']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.60
• base: 0.919
• confidence: 0.919 → use_as_is

2위 bim_dx_comparison_table (Frame 18) — reject

• anchor: 0.70 — bim_dx: ['BIM', 'DX'] (cap0.70 적용 (방증 0.29<0.5))
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.48
• base: 0.821
• confidence: 0.821 → reject

3위 app_sw_package_vs_solution (Frame 23) — reject

• anchor: 0.00 — None: []
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.47
• base: 0.645
• confidence: 0.645 → reject

4위 engn_sw_three_types (Frame 24) — reject

• anchor: 0.67 — engn_sw: ['Engn', 'S/W']
• cardinality: 0.50
• relation: 1.00
• slot: 0.50
• content: 0.50
• base: 0.641
• adaptation: -0.00 ['forbidden:n<min']
• confidence: 0.641 → reject

5위 bim_adoption_central_split (Frame 27) — restructure

• anchor: 0.67 — productivity_digital: ['DX', '디지털']
• cardinality: 1.00
• relation: 1.00
• slot: 1.00
• content: 0.61
• base: 0.838
• not_suits: -0.20 ['BIM vs DX 직접 대조']
• confidence: 0.638 → restructure

4. 요약

• 정답률: 4/4 — Phase 22~24 와 동일하지만 근거가 다름.
• 1-2위 격차: Phase 24 대비 일부 MDX에서 margin 작아짐 (2위도 구조 호환되면 light_edit 영역 상승).
• 하지만 운영 관점 차이:
  • 정답 모두 use_as_is (≥0.90) — 원본 그대로 slot fill.
  • 오답 후보 light_edit/restructure/reject 로 등급 분리 — 단순 1위/아님 이분법 아님.
  • 조건부 cap 의도대로 작동: MDX01-2 bim_dx 면제 (방증 comparison_dimensions 0.71), MDX03-2 bim_dx 적용 (방증 0.29<0.5).
  • not_suits 주체별 나열, 필수요건 등 구조 신호로 오답 후보 확실히 억제.

결론: template-fit-v1 은 단순 매칭 점수를 넘어 운영 가능한 DB 체계 제공. 기존 Phase 22~24 의 키워드/내용/구조 축은 base 의 일부로 유지하되, 감점 + 라우팅 + cap 이 실제 운영에 필요.