04 — dot-kg Process Evidence Graph (Workstream C)

Evidence drawn live from dot_tools (36 DOT_KG_*), v_axis_process_inventory, dot_iu_command_run, job_queue, event_type_registry, and the law §6/§7. No DOT executed.

4.1 Per-DOT evidence (all 36, by pair)

Common to all 36: status=active, execution_engine=pg_function, script_path/file_path NULL, last_executed NULL (0 runs), paired_dot set, component_count=1 in inventory. Input/output below are hypotheses from the law's process A–J text (no per-DOT spec exists).

pair	producer (Cấp B, on-demand, orphaned)	verifier (Cấp A, cron/dual, in inventory)	domain	proc	input → output (hypothesis)
1	EXTRACT	COMPLETENESS (cron)	kg.formation	A	row → entities+relations / audit coverage
2	CLASSIFY	VALIDATE (dual)	kg.formation	A	candidates → species+labels / validate→quarantine
3	LINK	CONSISTENCY (cron)	kg.formation	A	entities → edges+Qdrant / consistency audit
4	INTENT-CAPTURE	INTENT-RECALC (cron)	kg.business	B	chat/form → intent signal / recalc weights
5	RECOMMEND	FEEDBACK (cron)	kg.business	B	context → recommendation / feedback loop
6	JOURNEY-UPDATE	TIMELINESS (cron)	kg.business	B	event → journey node / staleness audit
7	CORRECT	ORPHAN (cron)	kg.quality	C	flag → correction / orphan-edge scan→auto-link
8	SIMILARITY	HEALTH (cron)	kg.quality	C/D	embedding → similar set / health snapshot 07:00
9	SCAFFOLD-BUILD	SCAFFOLD-VALIDATE (cron)	kg.scaffold	E	Đ38 → scaffold subgraph (dirty) / validate→clear dirty
10	SCOPE-PROJECT	SCOPE-VERIFY (cron)	kg.priority	F	person+role+goal → subgraph / ACL verify
11	PRIORITY-DECOMPOSE	PRIORITY-RECALC (cron)	kg.priority	F	goal → HTN priority tree / recalc on action done
12	DISCOVER+PROPOSE	CONSTRAINT-CHECK (cron)	kg.governance	G/C2	data → proposed edges (APR) / realtime compliance gate
13	PROVENANCE-TAG	PROVENANCE-AUDIT (cron)	kg.governance	gov	edge → provenance JSONB / audit "no-prov=quarantine"
14	EXPLAIN	EXPLAIN-VERIFY (dual)	kg.explain	C7	decision → explanation_path / verify path validity
15	SELF-SCORE	SELF-LEARN (dual)	kg.learning	I	outcome → self score / bounded weight update+snapshot
16	CONVERSE-EXTRACT	CONVERSE-VERIFY (cron)	kg.conversational	J	chat → extracted knowledge / auto-or-APR verify
17	EVICT-SCAN	EVICT-VERIFY (cron)	kg.eviction	H	stale scan → eviction candidates / verify→soft-delete
18	OVERRIDE-LOG	OVERRIDE-AUDIT (cron)	kg.governance	C12	override → log (who/what/why) / audit outcomes

Health/error signal: none available (0 executions, kg_quality_log not yet producing). Confidence per DOT: structural HIGH (defined, paired, governed, law-backed), runtime ZERO (never run).

4.2 Component graph

                 ┌──────────────── PRODUCERS (Cấp B, on-demand, event-fired) ────────────────┐
   Đ38 / data /  │  EXTRACT CLASSIFY LINK  INTENT-CAPTURE RECOMMEND JOURNEY-UPDATE  CORRECT   │
   chat / goal → │  SIMILARITY SCAFFOLD-BUILD SCOPE-PROJECT PRIORITY-DECOMPOSE               │ → universal_edges / KG state
                 │  DISCOVER+PROPOSE PROVENANCE-TAG EXPLAIN SELF-SCORE CONVERSE-EXTRACT       │
                 │  EVICT-SCAN OVERRIDE-LOG                                                   │
                 └───────────────────────────────┬───────────────────────────────────────────┘
                                                  │  each producer pairs 1:1 with…
                 ┌────────────────────────────────▼──────────────────────────────────────────┐
                 │  VERIFIERS (Cấp A, cron/dual, scheduled)  — "Cấp A IDLE = Cấp B đúng"        │
   cron 02–07h → │  COMPLETENESS VALIDATE CONSISTENCY INTENT-RECALC FEEDBACK TIMELINESS ORPHAN │ → kg_quality_log
                 │  HEALTH SCAFFOLD-VALIDATE SCOPE-VERIFY PRIORITY-RECALC CONSTRAINT-CHECK      │
                 │  PROVENANCE-AUDIT EXPLAIN-VERIFY SELF-LEARN CONVERSE-VERIFY EVICT-VERIFY     │
                 │  OVERRIDE-AUDIT                                                              │
                 └─────────────────────────────────────────────────────────────────────────────┘

The cluster is a continuous 24/7 loop (no single linear start→end): producers fire on events, verifiers audit on cron; HEALTH (07:00) snapshots the whole.

4.3 Suspected start / end nodes

• Start nodes (event-driven): EXTRACT (A, on PG insert/update), SCAFFOLD-BUILD (E, on Đ38 enact), INTENT-CAPTURE / CONVERSE-EXTRACT (B/J, on chat), DISCOVER+PROPOSE (G, weekly). There is no single start — the law defines 10 parallel processes, each with its own trigger.
• End nodes: none are terminal; HEALTH (D, 07:00 snapshot → kg_quality_log) is the closest to a "settle" point; SELF-LEARN (I) closes the feedback loop. The process is cyclic, not terminating — a key reason it cannot be modelled as one start→end pipeline.

4.4 Missing edges

• No materialised edges between DOTs. Pairing is the only relation (paired_dot); the producer→verifier and A→J sequencing exists only in the law text, not in any edge table (universal_edges has no DOT→DOT edges).
• No correlation edge: nothing links a producer run to its verifier run (no shared process_run_id).

4.5 Orphan components (the headline finding)

• 18 producer DOTs (on-demand) are orphaned from v_axis_process_inventory (which only ingests cron/event/dual/on-deploy). Surfaced live in v_process_discovery_orphan_components with orphan_class='ORPHAN_PRODUCER_PAIR_COVERED' — i.e. each one's verifier IS counted but the producer is not. This means the process axis literally shows half the KG automation.
• Inventory granularity orphan: even the 18 counted verifiers are component_count=1 singletons — the 18 pairs and the single A–J cluster are not represented as composite processes anywhere.

4.6 Duplicate / obsolete components

None obsolete. Possible near-duplicate by token outside KG (e.g. *-validate, *-verify across families) but within KG the 18 pairs are distinct. The law explicitly forbids splitting/renumbering the matrix (pair #12 silent-patch note).

4.7 Relation to IU 10649925

IU 10649925 (#7-36-dot-kg, technical_spec) is the sole content artifact describing these 36 DOTs. It is the textual bridge: topic knowledge_graph → this IU → 36 DOT_KG_* rows. The pairing matrix in the IU is reproduced 1:1 in dot_tools.paired_dot. No per-DOT IU exists below this one — D2 reconstructs operations from it + the process A–J text.

4.8 Scoreability

The graph is complete enough to score: nodes (36), pairs (18), roles (producer/verifier), triggers, domains, inventory-coverage, runtime-evidence (all zero), correlation (none for DOTs). The live v_process_discovery_quality_score consumes exactly this and returns dot:kg = strong_candidate_structural (confidence 50/100): full structure, zero runtime/correlation. See doc 08 for the scoring detail.