KB-1F54

D1 — Knowledge Graph Curated Cluster (Điều 39 consolidated) — DRAFT/CANDIDATE

10 min read Revision 1
knowledge-graphdieu39d1clustertechnical-designcandidate2026-06-04

D1 — Knowledge Graph: Curated Cluster (Điều 39 consolidated)

STATUS: DRAFT / CANDIDATE. This document curates the enacted text of dieu39-knowledge-graph-law.md into one navigable cluster. It is not itself law and does not promote, ratify, or alter anything. The underlying IUs are lifecycle_status=draft in the registry even though the law text is marked BAN HÀNH. The knowledge_graph topic is an AX-TOPIC CANDIDATE (no taxonomy node born). Authored 2026-06-04 by the architecture/governance agent from live evidence.

1. Purpose

Give Incomex one place to understand the Knowledge Graph as designed in Điều 39 v2.3 — its vision, definitions, architecture, the 10 operating processes, and the 36 dot-kg DOTs that implement them — and to connect that design to what is actually live in production today. Companion: D2 (d2-dot-kg-process-family-sop-runbook) covers the DOT family operationally.

2. Scope

In scope: the 16 IUs of dieu39-knowledge-graph-law.md; the 36 live DOT_KG_* DOTs; the 14 KG PG config tables; the topic/process Registries-Pivot axes. Out of scope: changing the law, executing any DOT, creating taxonomy nodes, registering processes.

3. Source IUs (the cluster)

16 sections of knowledge/dev/laws/dieu39-knowledge-graph-law.md (all law_unit, draft): #title, #0-tam-nhin, #1-26-bai-toan, #2-dinh-nghia, #3-4-nhom-a-b, #5-nhom-c, #6-10-quy-trinh, #7-36-dot-kg (IU 10649925, the bridge), #7b-bien-phap-quan-ly, #7c-moi-truong-thuc-thi, #8-kien-truc, #9-11-nt, #10-roadmap, #12-quan-he, #13-no-ky-thuat, #changelog. The knowledge_graph topic candidate tags 10 of these.

4. Vision & value chain (#0)

Mục tiêu: "Tạo ra Knowledge Graph đủ tin cậy để AI ra quyết định và điều hành hoạt động kinh doanh của Incomex."

Five-tier value chain: ① Data (PG) → ② Data Graph (edges, FK) → ③ Knowledge Graph (tri thức ngữ nghĩa) → ④ Priority Graph (ưu tiên theo context) → ⑤ AI quyết định → Con người giám sát.

Two data sources for the graph (S169): (1) entity_relations — soft/static relations (synonym/contains/contradiction) from Birth Registry Đ0-G; (2) Điều 38 output — live/dynamic bindings from policy/process/contract/chat documents. Nodes = born entities; static edges = entity_relations ∪ universal_edges (FK); dynamic edges = Đ38 bindings.

5. Definitions & trust model (#2)

Knowledge = information + context + inference. KG = network of entities + semantic relations. Triple (S)-[P]->(O). Ontology = Species (33, the SSOT) + Relation Types + Collection Schema + Labels. Semantic layer = Qdrant + PG metadata + ontology.

Trust survival gate (load-bearing): every edge + decision must carry confidence + freshness + provenance + source_authority. trust_score is computed only after every mandatory component clears its survival threshold in kg_thresholds. If any of confidence / freshness_decay / provenance_weight / source_authority is missing / invalid / zero-where-forbidden / below threshold → trust_score = 0 and the edge is quarantined. Only survivors get the product: trust_score = confidence × freshness_decay × provenance_weight × source_authority. Edge trust < threshold → suggestion only. Source authority order: quy định (Đ38) > báo cáo > chat (config kg_source_authority). "Không provenance = quarantine" (W3C PROV-O, problem A8).

6. The 26 problems & group C architecture (#1, #5)

26 problems in 4 groups: A infrastructure (8) — 3-tier data/graph/KG, ontology=Species, entity resolution, relation extraction, 5-dimension quality, self-healing, maturity model, provenance. B business (5) — employee/customer journeys, intent signal weighting, hyper-personalization, case-study transfer. C architecture (13, C1–C13). D operations (4) — DOT 100% (36 DOT/18 pairs), config-driven, agent-reads-KB, evolution tracking.

Group C highlights: C1 Scaffold (Đ38 → DOT-KG-SCAFFOLD-BUILD, PARTIAL rebuild on regulation change via scaffold_dependency_map, is_dirty flag), C2 Compliance Gate (DOT-KG-CONSTRAINT-CHECK → PASS/FAIL/WARNING; emergency mode is runtime config, stricter never looser), C5 Priority Graph (HTN 3-4 levels — hard boundary with Điều 34: Điều 39 emits only a flat static ordered task list [{task_id, primitive_task, target_state, evidence_refs}]; no if/else, retry, routing, state-machine — those belong to Điều 34), C9 Self-Learning (bounded ABS(new-old) ≤ max_delta, kg_weight_snapshots rollback, champion/challenger in kg_model_versions, PG trigger rejects ai_self_learn touching TBox), C13 Negative Knowledge (knowledge_type='negative', checked before any proposal).

7. Execution environment & architecture (#7c, #8)

Config-driven, zero hardcode (§8.2 — 9 worked examples, all "Code? ❌"): new signal = INSERT kg_signal_config; new constraint = INSERT kg_constraint_config; new DOT = register dot_tools (Đ35); etc. "Pipeline đọc PG MỖI LẦN CHẠY." The 14 KG PG tables: dot_tools, universal_edges, Qdrant entity_embeddings, kg_signal_config, kg_thresholds, kg_constraint_config, kg_acl_config, kg_auto_approve_rules, kg_source_authority, kg_priority_templates, scaffold_dependency_map, kg_weight_snapshots, kg_model_versions, kg_quality_log + kg_evolution_snapshots (*created only when C9 ships). Stack: PG16 (SSOT) + Qdrant (vectors), AGE phase 2-3. Sync (NT6, mandatory, no bypass): PG → Directus → Nuxt (screens) | Directus → AgentData (docs) | PG → Qdrant (vectors via SIMILARITY).

8. The 10 processes A–J & the 36 DOTs (#6-10, #7-36) — the bridge

Each process is dual-trigger (event/realtime + cron) and writes idempotently (deterministic key, OCC version/xmin, FOR UPDATE SKIP LOCKED; "Agent CẤM blind-update universal_edges").

proc name producer DOTs (Cấp B) verifier DOTs (Cấp A) cron
A Knowledge formation EXTRACT, CLASSIFY, LINK COMPLETENESS, VALIDATE, CONSISTENCY 02:00
B Business INTENT-CAPTURE, RECOMMEND, JOURNEY-UPDATE INTENT-RECALC, FEEDBACK, TIMELINESS 04:00
C Feedback CORRECT, SIMILARITY ORPHAN, HEALTH 05:00
D Monitoring (on-demand HEALTH red) COMPLETENESS, CONSISTENCY, TIMELINESS, ORPHAN, HEALTH 06:00–07:00
E Scaffold (partial invalidation) SCAFFOLD-BUILD SCAFFOLD-VALIDATE on Đ38 enact
F Priority decomposition SCOPE-PROJECT, PRIORITY-DECOMPOSE SCOPE-VERIFY, PRIORITY-RECALC daily
G Bottom-up discovery DISCOVER+PROPOSE CONSTRAINT-CHECK weekly
H Eviction EVICT-SCAN EVICT-VERIFY monthly
I Self-learning SELF-SCORE SELF-LEARN weekly
J Conversational CONVERSE-EXTRACT CONVERSE-VERIFY daily
(gov) Provenance / override PROVENANCE-TAG, OVERRIDE-LOG, EXPLAIN PROVENANCE-AUDIT, OVERRIDE-AUDIT, EXPLAIN-VERIFY cron

The 36 DOTs = 18 producer/verifier pairs; design principle "Cấp A IDLE = Cấp B đúng = thiết kế tốt" (an idle verifier means the producer is correct). Runtime rule for pair #12: CONSTRAINT-CHECK must run as a realtime gate, not wait for weekly DISCOVER+PROPOSE. Full per-DOT operations → D2.

9. Governance & the golden rule (#7b, Đ37)

TBox (schema / scaffold / new species / ontology) → human approval via Council / Data Stewards (Đ32 APR, Đ37). ABox-large (override rules) → department management. ABox-routine (link / weight / intent) → AI auto with guardrails. "Cơ quan chủ quản KG phải đăng ký governance_registry (Đ37). Chưa đăng ký = chưa triển khai."the KG owner is not yet registered; this is the central governance blocker.

10. Relation to Registries-Pivot

  • AX-TOPIC (CANDIDATE): knowledge_graph is a topic candidate (10 IUs), the convergence hub; sibling architecture (5 IUs, 3 shared).
  • AX-PROCESS (CANDIDATE): the dot-kg DOTs are Type-1 DOT-contained components. Caveat (see D2 / discovery report): v_axis_process_inventory currently sees only the 18 verifier DOTs; the 18 producers are orphaned.
  • Joint map: v_information_piece_topic_process_map ties knowledge_graph topic ↔ KG_DOT_CLUSTER (relation SHARED_IU+DOMAIN).

11. Current gaps

  1. No per-DOT spec content (one matrix IU only). 2. No KG config-table reference doc. 3. KG owner unregistered (governance_registry). 4. Topic→knowledge_documents bridge unpopulated. 5. Runtime empty: 0 DOT executions, 0 KG events — the design has never run. 6. IU lifecycle draft vs header BAN HÀNH.

12. Next actions

  • Author the 14 config-table reference (D-series follow-on).
  • Register the KG owner in governance_registry (owner action, Đ37).
  • Decide promotion of the knowledge_graph topic to a FAC-08 node (AXR-TOPIC-03/14, owner-gated).
  • Populate the topic→document bridge once a node exists.
  • Use the live discovery views to track DOT runtime/correlation when execution begins.

13. Evidence references

dieu39-knowledge-graph-law.md (16 IUs); dot_tools (36 DOT_KG_*, 18 pairs, 0 executions); v_information_piece_topic_process_map; v_axis_process_inventory; the 6 v_process_discovery_* views; checkpoints checkpoint-information-piece-topic-process-content-work-2026-06-04, checkpoint-process-workflow-axis-rp-pilot-2026-06-04. Live-verified 2026-06-04; 0 production mutation by this document.