Ontology, Knowledge Graph, and Semantic Data Intake Virtuoso — Boss Agent and MOA Registration

Version: 1.0.0 Registered: August 6, 2026 Status: Code-side registration complete on the feature branch; production activation follows the normal Boss/MOA/Virtuoso deployment pipeline. Do not claim production registration until the serving-plane checks at the bottom pass.

Purpose

One additive, governed semantic-governance skill — ontology-kg-virtuoso — giving Boss Agent and MOA eight routable expertise profiles (a parent orchestrator plus seven specialists) across ontology engineering, ontology research, controlled ontology evolution, semantic data intake, entity resolution, knowledge-graph quality, and graph-powered retrieval/reasoning. It is the canonical owner of semantic governance: canonical concepts and definitions, source mappings, identity semantics, provenance/temporal/contradiction rules, ontology versions, and graph integrity. It extends miz-oki-platform-expert and complements enterprise-virtuoso (whose knowledge_acquisition_virtuoso discovers and trust-scores the sources this skill semantically governs) and media-acquisition-virtuoso; it replaces neither and removes nothing. It also establishes the repository's first canonical ontology registry (ontology/).

What is registered

The profile names are skill overlays, not model IDs. Model selection remains governed by the central Virtuoso role registry (roles used: CODING_ARCH, DATA_CAUSAL). Every profile is recommend_only.

Registration chain

skills/ontology-kg-virtuoso/                     (CANONICAL — 17 files)
├── SKILL.md
└── references/
    ├── profile_registry.yaml                    (8 profiles, machine-readable)
    ├── routing_matrix.yaml                      (routing rules + smoke routes)
    ├── output_contracts.md                      (28-item decision package + claim labels)
    ├── governance_contract.md                   (17 additive hard rules)
    ├── tool_capability_map.yaml                 (capability CATEGORIES per profile)
    ├── semantic_intake_contract.md              (mandatory 20-step intake + 9 mapping statuses + quarantine)
    ├── ontology_evolution_contract.md           (lifecycle + proposal template + 3 approval levels)
    ├── provenance_temporal_contradictions.md    (provenance fields, 9 assertion types, temporal axes, contradiction protocol)
    ├── quality_metrics.yaml                     (metric DEFINITIONS — design targets, not results)
    └── domain_playbooks/*.md                    (7 specialist doctrines)
        │
        ├─ byte-identical mirror → .claude/skills/ontology-kg-virtuoso/            (Claude Code discovery)
        ├─ byte-identical mirror → src/shared/virtuoso_models/skills_data/ontology_kg_virtuoso/
        └─ byte-identical mirror → services/virtuoso-models-service/virtuoso_models/skills_data/ontology_kg_virtuoso/

skills/miz-oki-platform-expert/references/boss_agent_skillpack.md
└── §10 ▸ "### Ontology & Knowledge Graph Virtuoso Registry v1.0 — Boss + MOA"   (the runtime registration)
    ├─ Boss startup: skillpack_bootstrap.compose_skillpack_context()  → system prompt (full mode)
    ├─ MOA experts: model_factory._moa_system_instruction()           → shared skill context
    ├─ every virtuoso_call(..., skills=True)                          → role-sliced context
    └─ MCP: skills_get_section("10")                                  → section retrieval

miz-oki-adk-agents/config/config.yaml ▸ agents.specialists
└── 8 ontology profiles: model_role + provider + skill_profile + skill_source
    + execution_posture: "recommend_only" (Acquisition's 5 and Enterprise's 13 remain above, unchanged)

mcp/mizoki-skills-mcp/ (additive extension)
├── ontology_core.py                             (transport-free, stdlib-only)
└── server.py registers:
    ├─ resource mizoki://skills/ontology-kg-virtuoso               (SKILL.md verbatim)
    ├─ resource mizoki://skills/ontology-kg-virtuoso/profiles      (registry YAML)
    ├─ resource mizoki://skills/ontology-kg-virtuoso/routing       (routing matrix YAML)
    ├─ resource mizoki://skills/ontology-kg-virtuoso/capabilities  (capability map YAML)
    ├─ tool ontology_skills_list_profiles
    ├─ tool ontology_skills_get_profile
    ├─ tool ontology_skills_route
    ├─ tool ontology_skills_get_capabilities
    └─ tool ontology_skills_version

ontology/                                        (NEW — canonical ontology registry, first in repo)
├── registry.yaml (version 1.0.0, documented baseline; inventories existing schema homes)
├── foundational/ (33 classes, 28 relationships incl. OPERATING_SYSTEM §2.2 set, 18 constraints)
├── enterprise/ (36 business concepts)
├── mappings/ + competency-questions/ + proposals/ + migrations/ + changelog/

src/cells/cell06/workflows/miz3-semantic-intake-v1.json   (Cell 6 MOE-Router workflow manifest,
    additive; discoverable via GET /workflows; existing identity-resolution manifest unchanged)

The skillpack §9 GOVERNANCE section is untouched — its SHA-256 (6c529ac9e60a2e8c7079058f51d232a14f4d431185d75b0d6b949bfd1aaa4eee) and the 12-entry section map are unchanged, so no fingerprint pin churn occurred.

Routing behavior

• Single-scope requests route to the matching specialist.
• Requests spanning ≥2 semantic scopes route first to ontology_kg_virtuoso, which builds the delegation plan and dispatches specialists to MOA.
• ontology_evolution_virtuoso auto-includes on canonical semantic changes (class merge/split, identifier/definition change, constraint change, deprecation); entity_resolution_virtuoso on identity/merge/person-household work; graph_reasoning_virtuoso on retrieval/reasoning impact measurement.
• Cross-skill: knowledge_acquisition_virtuoso (Enterprise) joins for source discovery/trust scoring; legal_compliance_virtuoso for external-ontology licensing and sensitive-data authorization; the Media Acquisition profiles for marketing/commerce source onboarding. Existing routes of both prior skills are unchanged; when multiple skills apply, all are used.
• MOA specialists analyze independently, cite evidence, state uncertainty, and preserve dissent; consensus can never override a failed validation gate.

Governed operating boundary

Evidence pipeline: governed connectors → service-canonical-ingestion → canonical event envelope (MAPPERS[source](raw) → ingest_gate) → KG projection. No direct writes to the canonical KG, BigQuery, Firestore, or vector indexes; semantic validation completes before canonical integration, and data that cannot pass is quarantined with reason codes, never forced or silently dropped. Mappings carry the nine-status vocabulary projecting onto the envelope's coarse mapping_status (the shape registry always wins). Identity is deterministic-first, consent-gated, confidence-scored, reversible; probabilistic identity/household edges stay recall-only and out of causal measurement; raw email/phone never enter the KG. Every assertion carries provenance, confidence, assertion type, and temporal metadata; predictions/recommendations/opinions/hypotheses are never served as observed facts; history is superseded, never overwritten; contradictions are preserved with recorded-evidence-only resolution. Ontology changes are versioned, audited migrations under three approval levels (automatic/steward/architecture) with no self-approval above automatic; irreversible changes are permanently approval-gated. No hallucinated standards, no unlicensed external ontology adoption, no mock benchmark results; improvement claims require measured before/after benchmarks.

Verification

python scripts/ontology_skills_sync.py --check --json    # 8 profiles, mirrors, governance, MCP, ontology/ (incl. loader validation battery), legacy intact
python scripts/skills_sync.py --check                    # base Skills v3.0 parity (incl. vscode extension bundled copy)
python scripts/acquisition_skills_sync.py --check --json # Media Acquisition unchanged
python scripts/enterprise_skills_sync.py --check --json  # Enterprise unchanged
python3 src/shared/ontology_registry.py --check          # governed registry loader: zero violations
python3 scripts/ontology_eval.py --check                 # benchmark artifact byte-reproducible
python -m pytest -q tests/skills/test_ontology_kg_virtuoso.py
python -m pytest -q tests/skills                         # full skills suite

Expected: Ontology-KG Virtuoso parity: OK (8 profiles; governance preserved; legacy skills intact). Measured on the feature branch (2026-08-06): all four gates OK; tests/skills 99 passed, 1 skipped (pre-existing MCP-SDK-dependent skip). CI: .github/workflows/ontology-kg-virtuoso-skills-parity.yml (PR + push to main/agent/** + dispatch); the umbrella ci.yaml skills gate also runs the test file via pytest tests/skills -q.

Production rollout

1. Merge to main through the repository's established workflow (AI-branch auto-merge or reviewed PR).
2. The Deploy Router matches the changed paths: src/shared/** and miz-oki-adk-agents/config/config.yaml → deploy-boss-agent-core.yml; src/shared/virtuoso_models/** → deploy-coding-moa.yml; services/virtuoso-models-service/** → deploy-virtuoso-models.yml. The MOA controller (deploy-moa-controller.yml) watches only miz-oki-adk-agents/moa/** and must be dispatched manually so its vendored skillpack hash matches the Boss.
3. Post-deploy serving-plane checks (all must pass before claiming production registration):

# Boss serves the registry (all 8 profile ids + the marker):
curl -s -X POST "$BOSS_URL/execute" -H "Content-Type: application/json" \
  -d '{"tool":"skills_get_section","section_name":"10"}'
# Skillpack context loaded, full mode, hash reported:
curl -s -X POST "$BOSS_URL/execute" -H "Content-Type: application/json" \
  -d '{"tool":"skills_version"}'
# latestReady == latestCreated on boss-agent-adk, coding-moa, virtuoso-models-service,
# miz-oki-moa-controller; unauth /health posture unchanged (403 = IAM-locked, healthy).

Rollback

• Code: revert the Ontology-KG Virtuoso commits (all changes are additive; reverting restores the prior skillpack/config byte-exactly), rerun all four parity gates.
• Runtime: route traffic to the prior known-good Cloud Run revision per service; verify health and the prior skillpack hash.
• Feature: the existing global kill switch VIRTUOSO_SKILLS_DISABLED=1 disables all skillpack context (base + acquisition + enterprise + ontology) on every surface; ENABLE_SKILLPACK_CONTEXT=false disables the Boss bootstrap only. No ontology-specific flag was added: the registration is inert data unless the shared skillpack layer serves it, and a scoped kill switch would have required new code in the hash-pinned load path. Removing the §10 H3 block (a one-commit revert) is the scoped disable.

Known limitations

• The registration lives in skillpack §10: DATA_CAUSAL/CREATIVE_MM role slices of virtuoso_call do not include §10 (same as the acquisition and enterprise skills); Boss (full mode), MOA experts (role=None), CODING_ARCH, and DEVOPS_OPS receive it, and any agent can pull it via skills_get_section("10").
• ontology_skills_route is deterministic keyword routing over registry aliases — advisory metadata for Boss, not a semantic classifier.
• Capability categories map to live tools only through runtime discovery; unconfigured stores report not_configured, never mock data.
• Specialist profiles are governed config entries composed on demand (same as acquisition/enterprise); they are not eight always-on services.
• ontology/ is a machine-validated baseline (spec authority, registry v1.1.0 / OCP-20260806-001): the governed loader src/shared/ontology_registry.py parses and structurally validates every module (reference integrity, subclass acyclicity, named approval-ladder appointments, evaluation-set integrity, content hash), the mizoki-skills MCP server consumes it at runtime (ontology_registry_get/ontology_registry_validate + mizoki://ontology/registry), and CI fails on any violation via scripts/ontology_skills_sync.py --check. Per-service ADOPTION of ontology-governed behavior still lands through governed proposals/migrations — the loader existing opts no service in. The envelope's ontology_version/mapping_status fields and the existing schema homes (kg_schemas.py, UEES, kg_ingestion_config.yaml, mizoki_finance graph labels, marketing_ontology.yaml) remain the serving surfaces and are inventoried, not modified.
• The steward/architecture approval rungs are operational: named appointments live in ontology/registry.yaml (approval_ladder_appointments — owner directive 2026-08-06: the Owner, ceo@mediaintelligence.ai, on both rungs; delegation per GOVERNANCE.md 1.1.4), and the loader fails CI when a rung is empty. Approval identity at decision time still derives from authentication.
• Retrieval/reasoning metrics have governed evaluation sets (ontology/evaluation/, frozen v1.0.0) with a deterministic harness (scripts/ontology_eval.py) and a committed byte-reproducible baseline artifact (evaluation/results/structural-baseline-v1.json). Scope honesty: the baseline measures the documented registry (lexical baseline retriever + structural reasoning); live GraphRAG/CausalRAG figures remain design targets until a live evaluation artifact exists (acceptance_eligible_for_live_claims: false).
• The Cell 6 manifest makes the semantic-intake workflow discoverable (GET /workflows); step execution binds to services incrementally — the manifest is declarative registration, not a claim that all 20 steps are automated today.
• The vscode-boss-agent-extension/resources/boss_agent_skillpack.md bundled copy is re-synced byte-identical to canonical and standing-gated (extension v3.0.2, 2026-08-06): scripts/skills_sync.py --check now verifies it and --fix re-syncs it, so it can no longer drift silently as it did through the Acquisition/Enterprise generations. Repackaging/installing the .vsix remains an operator-local step (npx @vscode/vsce package --no-dependencies).