diff --git a/rag-agentic-dashboard/data/enterprise-aigov-framework.json b/rag-agentic-dashboard/data/enterprise-aigov-framework.json new file mode 100644 index 0000000..6ab9a15 --- /dev/null +++ b/rag-agentic-dashboard/data/enterprise-aigov-framework.json @@ -0,0 +1,2999 @@ +{ + "docRef": "ENTERPRISE-AIGOV-FRAMEWORK-WP-058", + "version": "1.0.0", + "title": "Enterprise AI/AGI Governance Framework for Large Financial & Fortune 500 Enterprises (2026-2030)", + "horizon": "2026-2030", + "apiPrefix": "/api/enterprise-aigov-framework", + "buildsOn": [ + "WP-035", + "WP-040", + "WP-045", + "WP-050", + "WP-054", + "WP-055", + "WP-056", + "WP-057" + ], + "status": "regulator-submission-grade", + "classification": "Confidential / Restricted \u2014 Board, CRO, CCO, CISO, CDAO, Group Internal Audit, External Regulators (on request)", + "directive": { + "scope": "End-to-end enterprise AI/AGI governance operating model for Fortune 500 / Global 2000 / G-SIFIs spanning policy, control, risk, compliance, security, model risk, third-party, AGI containment, and AI Governance Hub architecture", + "outcomes": [ + "ISO/IEC 42001:2023 certified AIMS by 2027 across all material AI systems", + "NIST AI RMF 1.0 + AI 600-1 generative profile mapped to >=95% of in-scope models", + "EU AI Act 2026 Article 6/9/10/14/15 + GPAI 53/55 obligations fully evidenced", + "GDPR DPIA + Article 22 + FCRA/ECOA adverse-action automation in production", + "Basel III/IV + SR 11-7 + OCC 2011-12 model risk lifecycle managed in single MRM platform", + "FCA Consumer Duty + SMCR SMF-attested AI accountability operating model", + "MAS FEAT + HKMA GP-1/GS-2 fairness, ethics, accountability, transparency in production", + "Kafka audit log with WORM + PQC sealing across all model decisions", + "Kubernetes + container security with policy-as-code (OPA/Rego) at admission and runtime", + "AGI containment T0-T4 with 3-of-5 quorum + kinetic override and AI Safety Institute notifications" + ], + "doNot": [ + "Do NOT deploy any AI/AGI capability without Enterprise AI Governance Hub registration, ISO 42001 risk assessment, and model risk tier classification", + "Do NOT bypass Kafka audit logging, OPA/Rego policy gates, WORM/PQC sealing, or 3-of-5 frontier quorum" + ] + }, + "regimes": [ + "ISO/IEC 42001:2023 AIMS", + "ISO/IEC 23894:2023 AI Risk", + "ISO/IEC 27001:2022 ISMS", + "ISO/IEC 27701:2019 PIMS", + "NIST AI RMF 1.0", + "NIST AI 600-1 Generative AI Profile", + "NIST SP 800-53 Rev.5", + "NIST SP 800-218 SSDF", + "OECD AI Principles 2019/2024", + "EU AI Act 2024/1689", + "EU AI Act GPAI Art. 53/55", + "EU GDPR + Art. 22", + "EU DORA", + "EU NIS2", + "EU CRA", + "US FCRA + ECOA Reg-B", + "US Fed SR 11-7 + OCC 2011-12", + "Basel III/IV + ICAAP", + "US SEC 17a-4 + 10-K/8-K", + "FINRA 3110/4511", + "UK FCA Consumer Duty", + "UK SMCR + SS1/23", + "MAS FEAT + TRM 2021", + "HKMA GP-1 + GS-2", + "OSFI E-23", + "FINMA AI Guidance", + "G7 Hiroshima Process", + "Bletchley/Seoul/Paris Declarations" + ], + "indices": { + "AIMS-Coverage": ">=0.95 (ISO 42001 controls)", + "MRGI": ">=0.95 (Model Risk Governance Index, SR 11-7 + OCC 2011-12)", + "DRI": ">=0.95 (Decision Reproducibility, n=10)", + "CCS": ">=0.95 (Control Coverage Score)", + "ARI": ">=0.9 (Alignment Robustness Index, frontier)", + "CSI": ">=0.95 (Containment Sufficiency, T3/T4)", + "RTRI": ">=0.9 (Red-Team Resilience Index)", + "CDC-Score": ">=0.9 (FCA Consumer Duty compliance)", + "RCI": "=1.0 (Regulator Confidence Index)" + }, + "tiers": { + "T0": "Sandbox - isolated VPC, synthetic data only, no production traffic", + "T1": "Staging - shadow mode, real data, no customer impact", + "T2": "Canary - <=1% production traffic, automated rollback", + "T3": "Production - Nitro Enclaves + KMS + dual control + full audit", + "T4": "Frontier Air-Gapped - 3-of-5 quorum + kinetic override + AI Safety Institute notification" + }, + "severities": { + "SEV-0": "Civilizational/systemic - EU AI Office notification <=15d, AISI notification <=24h", + "SEV-1": "Major - SEC 8-K <=4 BD, DORA <=4h, FCA <=72h", + "SEV-2": "Material - regulator notification <=72h", + "SEV-3": "Operational - internal escalation <=10 BD" + }, + "investment": { + "envelope": "USD 180-500M / 5y (Fortune 500 / G-SIFI tier)", + "NPV": "USD 500-1500M (5y, risk-adjusted)", + "drivers": [ + "MRM platform consolidation", + "AI Governance Hub build", + "Kafka audit + WORM/PQC sealing", + "Kubernetes + OPA/Rego at scale", + "AGI containment T3/T4 enclaves", + "Red-teaming program", + "Regulator attestation tooling" + ] + }, + "counts": { + "modules": 9, + "sections": 45, + "policies": 15, + "controls": 25, + "kafkaTopics": 12, + "k8sControls": 15, + "opaPolicies": 15, + "wormControls": 12, + "mrmArtifacts": 15, + "redTeams": 15, + "agiContainments": 15, + "hubComponents": 16, + "schemas": 16, + "code": 15, + "kpis": 30, + "riskControlMatrix": 16, + "traceability": 20, + "dataFlows": 12, + "regulators": 16, + "rollout90": 3, + "roadmap": 5, + "evidencePack": 16 + }, + "modules": [ + { + "mid": "M1", + "title": "ISO/IEC 42001 AIMS + NIST AI RMF + OECD + EU AI Act Foundation", + "summary": "Integrated AI Management System anchored on ISO/IEC 42001:2023 with NIST AI RMF 1.0 functions (Govern/Map/Measure/Manage), OECD AI Principles, and EU AI Act 2024/1689 Article 6/9/10/14/15 + GPAI 53/55 mappings.", + "sections": [ + { + "sid": "M1.1", + "title": "ISO/IEC 42001 AIMS Architecture", + "policyCommit": "Board-attested AI Policy, AI Objectives, AI Risk Appetite Statement signed annually by Group CEO + Group CRO", + "scope": "All AI/AGI systems with material impact on customers, employees, regulators, capital, or systemic risk", + "structure": "AIMS clauses 4-10 mapped to enterprise functions: Context (Strategy), Leadership (CDAO+CRO), Planning (PMO), Support (HR+Procurement+IT), Operation (Lines of Business), Performance (Internal Audit), Improvement (CCO)" + }, + { + "sid": "M1.2", + "title": "NIST AI RMF 1.0 + AI 600-1 Generative Profile", + "functions": [ + "GOVERN", + "MAP", + "MEASURE", + "MANAGE" + ], + "generativeProfile": "NIST AI 600-1 mapped to all FM/LLM use cases with synthetic content provenance (C2PA)", + "crossWalk": "Bidirectional crosswalk ISO 42001 <-> NIST AI RMF <-> EU AI Act maintained in MRM platform" + }, + { + "sid": "M1.3", + "title": "OECD AI Principles 2019/2024 Implementation", + "principles": [ + "Inclusive growth", + "Human-centered values", + "Transparency", + "Robustness", + "Accountability" + ], + "implementation": "OECD-aligned model cards + system cards published for all T2+ systems with public-facing summary" + }, + { + "sid": "M1.4", + "title": "EU AI Act 2024/1689 Compliance Layer", + "riskClasses": [ + "Unacceptable (Art. 5)", + "High-risk (Art. 6/Annex III)", + "Limited-risk (Art. 50)", + "Minimal-risk" + ], + "highRiskObligations": [ + "Art. 9 risk mgmt", + "Art. 10 data governance", + "Art. 14 human oversight", + "Art. 15 accuracy/robustness/cybersecurity" + ], + "gpai": [ + "Art. 53 technical documentation + copyright policy", + "Art. 55 systemic-risk: evaluations, adversarial testing, cybersecurity, incident reporting" + ], + "timeline": "Prohibited practices Feb 2025; GPAI Aug 2025; high-risk Aug 2026" + }, + { + "sid": "M1.5", + "title": "Board + Executive Accountability", + "committees": [ + "Board AI Risk Committee (quarterly)", + "Executive AI Governance Committee (monthly)", + "AI Ethics Council (monthly)", + "Model Risk Committee (weekly)" + ], + "roles": { + "AI-SMF (SMF-AI)": "FCA-attested senior manager", + "CDAO": "Operating accountability", + "CRO": "Risk appetite owner", + "CCO": "Regulatory engagement", + "CISO": "Cybersecurity + integrity", + "GIA": "Independent assurance" + } + } + ] + }, + { + "mid": "M2", + "title": "Financial-Services Model Risk Management (SR 11-7 / OCC 2011-12 / Basel III/IV / ICAAP)", + "summary": "Three-lines-of-defense model risk operating model with SR 11-7 conceptual soundness, ongoing monitoring, outcomes analysis; OCC 2011-12 effective challenge; Basel III/IV IRB/IMA model validation; ICAAP integration with Pillar 2.", + "sections": [ + { + "sid": "M2.1", + "title": "Model Risk Lifecycle", + "stages": [ + "Identification", + "Development", + "Validation", + "Approval", + "Implementation", + "Monitoring", + "Retirement" + ], + "tiering": "Tier-1 (regulatory capital, P&L, capital plan) / Tier-2 (material business) / Tier-3 (limited scope) / Tier-4 (research)", + "cadence": "Tier-1 annual validation; Tier-2 biennial; Tier-3 every 3y; ongoing monitoring monthly for all" + }, + { + "sid": "M2.2", + "title": "SR 11-7 + OCC 2011-12 Effective Challenge", + "conceptualSoundness": "Independent review of theory, assumptions, design choices", + "ongoingMonitoring": [ + "Backtesting", + "Benchmarking", + "Sensitivity analysis", + "Stress testing" + ], + "outcomesAnalysis": "Champion/challenger + counterfactual analysis on production decisions" + }, + { + "sid": "M2.3", + "title": "Basel III/IV IRB/IMA + FRTB", + "scope": [ + "PD/LGD/EAD IRB models", + "VaR/ES IMA models (FRTB)", + "AMA op-risk (legacy)", + "CCAR/DFAST stress models", + "IFRS 9/CECL ECL models" + ], + "validation": "Independent validation per SR 15-19/SR 15-18; quantitative review every cycle", + "capitalImpact": "MRM platform feeds Pillar 2 model risk capital add-on into ICAAP" + }, + { + "sid": "M2.4", + "title": "AI/ML-Specific MRM Extensions", + "extensions": [ + "Concept drift + data drift monitoring (PSI, KS, KL, Wasserstein)", + "Fairness across protected classes (FCRA/ECOA aligned)", + "Explainability evidence (SHAP/LIME/integrated gradients) per decision", + "Adversarial robustness testing (PGD, BIM, NLP attacks)", + "Provenance of training data + lineage to feature store" + ] + }, + { + "sid": "M2.5", + "title": "ICAAP / Pillar 2 Integration", + "integration": "Aggregate model risk capital + AI-specific add-ons fed into ICAAP Pillar 2 alongside operational, reputational, strategic risk", + "governance": "MRC quarterly review of capital adequacy; Board-attested annual ICAAP includes AI risk section" + } + ] + }, + { + "mid": "M3", + "title": "Data Protection & Consumer Fairness (GDPR / FCRA / ECOA / FCA Consumer Duty)", + "summary": "Privacy-by-design with GDPR Article 22 (automated decisions), FCRA adverse-action notices, ECOA Reg-B disparate impact testing, FCA Consumer Duty cross-cutting rules + four outcomes; MAS FEAT + HKMA GP-1 fairness operationalization.", + "sections": [ + { + "sid": "M3.1", + "title": "GDPR Article 22 + DPIA Operationalization", + "dpia": "DPIA required for all T2+ models processing personal data; reviewed by DPO + Group Legal", + "article22": { + "prohibition": "No solely automated decisions producing legal/significant effects without explicit consent or necessity", + "rights": [ + "Human intervention", + "Express view", + "Contest decision" + ], + "logging": "All Art-22 invocations logged to Kafka audit topic" + }, + "crossBorder": "EU SCC + UK IDTA + adequacy assessments documented in ROPA" + }, + { + "sid": "M3.2", + "title": "FCRA + ECOA Reg-B Adverse-Action Automation", + "adverseAction": "Automated FCRA 615(a) + ECOA Reg-B section 1002.9 notices generated within 30 days of adverse decision", + "reasonCodes": "Top-N reason codes derived from SHAP attribution, mapped to plain-English statements vetted by Compliance Legal", + "disparateImpact": "Quarterly disparate-impact analysis on credit, hiring, insurance models (race, sex, age, national origin)" + }, + { + "sid": "M3.3", + "title": "FCA Consumer Duty + Cross-Cutting Rules", + "outcomes": [ + "Products & services", + "Price & value", + "Consumer understanding", + "Consumer support" + ], + "crossCutting": [ + "Act in good faith", + "Avoid foreseeable harm", + "Enable customers to pursue financial objectives" + ], + "evidence": "Consumer Duty Board Report annual; AI-driven personalization included with foreseeable-harm assessment", + "vulnerableCustomers": "Vulnerable-customer flag piped to AI systems; fairness uplift required where applicable" + }, + { + "sid": "M3.4", + "title": "MAS FEAT + HKMA GP-1 / GS-2", + "feat": [ + "Fairness", + "Ethics", + "Accountability", + "Transparency" + ], + "hkmaGP1": "Governance of AI applications in banking \u2014 board-level accountability, risk management, fair treatment", + "hkmaGS2": "Generative AI applications guidance \u2014 data governance, model governance, human oversight, cybersecurity" + }, + { + "sid": "M3.5", + "title": "Privacy-Enhancing Technologies (PETs)", + "pets": [ + "Differential privacy (epsilon budgets per dataset)", + "Federated learning with secure aggregation", + "Homomorphic encryption (CKKS/BGV)", + "Secure multi-party computation", + "Confidential computing (AMD SEV-SNP, Intel TDX, AWS Nitro)" + ], + "policy": "PETs mandated for cross-border training and any T3 model handling special category data" + } + ] + }, + { + "mid": "M4", + "title": "Kafka-Based Audit Logging + WORM Storage + Post-Quantum Cryptography", + "summary": "Enterprise-wide tamper-evident audit log over Apache Kafka with WORM (S3 Object Lock COMPLIANCE / Azure Immutable / GCS Bucket Lock) and PQC-signed seals (ML-DSA / ML-KEM / SLH-DSA) per NIST FIPS 203/204/205.", + "sections": [ + { + "sid": "M4.1", + "title": "Kafka Audit Topic Architecture", + "topics": [ + "aigov.decisions", + "aigov.policy-changes", + "aigov.model-lifecycle", + "aigov.access", + "aigov.containment-events", + "aigov.regulator-notifications" + ], + "retention": "Hot 90d in Kafka tiered storage; cold WORM 7-25y per regime (SEC 17a-4 7y, GDPR varies, FINRA 6y, DORA 5y)", + "partitioning": "By LOB + decisionId; replication factor 3 across AZs; minISR=2; producer acks=all" + }, + { + "sid": "M4.2", + "title": "Tamper-Evident Sealing", + "chain": "Each record hashed (SHA-3-512); merkle-tree aggregated per minute; root signed with ML-DSA-87 (FIPS 204) + SLH-DSA fallback", + "anchoring": "Daily merkle root anchored to QLDB + external timestamp authority (TSA RFC 3161) + optional public chain", + "verification": "Independent verifier service replays Kafka offsets and recomputes merkle roots on demand" + }, + { + "sid": "M4.3", + "title": "WORM Storage Tier", + "backends": [ + "AWS S3 Object Lock COMPLIANCE mode", + "Azure Blob immutable storage policy", + "GCS Bucket Lock", + "On-prem Dell ECS / NetApp SnapLock Compliance" + ], + "policy": "Legal-hold + retention-lock dual control; no delete path even for root accounts; SEC 17a-4(f) WORM attestation by independent third party" + }, + { + "sid": "M4.4", + "title": "Post-Quantum Cryptography Stack", + "algorithms": [ + "ML-KEM-1024 (FIPS 203) for key encapsulation", + "ML-DSA-87 (FIPS 204) for signatures", + "SLH-DSA-SHA2-256s (FIPS 205) as conservative fallback" + ], + "hybrid": "Hybrid TLS classical+PQ during transition (X25519+ML-KEM-768) per NSA CNSA 2.0 timeline 2025-2033", + "keyMgmt": "HSM-backed (AWS CloudHSM / Azure Dedicated HSM / Thales Luna 7) with FIPS 140-3 Level 3" + }, + { + "sid": "M4.5", + "title": "Audit Query + Regulator Access", + "queryLayer": "ksqlDB + Trino over Iceberg on WORM; row-level filters per LOB + regulator scope", + "regulatorPortal": "Read-only portal for EU AI Office, FCA, MAS, HKMA, SEC, FINRA with audit-of-audit logging", + "sla": "Regulator query response <=24h; bulk export <=72h" + } + ] + }, + { + "mid": "M5", + "title": "Container & Kubernetes Security for AI Workloads", + "summary": "Defense-in-depth for AI model serving on Kubernetes spanning image supply chain (SLSA L4), admission control, runtime security, network policy, secrets, and confidential containers.", + "sections": [ + { + "sid": "M5.1", + "title": "Image Supply Chain (SLSA L4 / SSDF)", + "controls": [ + "Cosign signatures on all images", + "SBOM (SPDX/CycloneDX) per image", + "Vulnerability scanning (Trivy/Snyk/Prisma) in CI", + "Provenance attestations (in-toto)", + "Sigstore Rekor transparency log" + ], + "policyGate": "Kyverno/OPA admission rejects unsigned, unscanned, or non-policy-compliant images" + }, + { + "sid": "M5.2", + "title": "Admission Control + Pod Security", + "psa": "Pod Security Admission 'restricted' profile cluster-wide for AI namespaces", + "policyEngines": [ + "Kyverno", + "OPA Gatekeeper", + "Validating admission policies (VAP)" + ], + "controls": [ + "No privileged", + "No host network/PID/IPC", + "Read-only root FS", + "Non-root UID", + "Seccomp RuntimeDefault", + "Drop ALL capabilities" + ] + }, + { + "sid": "M5.3", + "title": "Runtime Security", + "tools": [ + "Falco for syscall anomaly detection", + "Tetragon eBPF for kernel-level enforcement", + "Cilium for network policy + observability" + ], + "response": "Auto-isolation + Kafka aigov.containment-events on anomaly; SRE+SecOps page within 5 min" + }, + { + "sid": "M5.4", + "title": "Network Policy + Service Mesh", + "netpol": "Default-deny Cilium NetworkPolicy + L7 HTTP/gRPC filtering", + "mesh": "Istio/Linkerd mTLS for service-to-service; SPIFFE/SPIRE workload identities", + "egress": "Per-namespace egress allowlist with FQDN policies; DNS over HTTPS" + }, + { + "sid": "M5.5", + "title": "Confidential Containers + Secrets", + "confidential": "Confidential containers (CoCo) on AMD SEV-SNP / Intel TDX / AWS Nitro Enclaves for T3/T4 workloads", + "secrets": [ + "Vault (HashiCorp) with auto-rotation", + "AWS Secrets Manager / Azure Key Vault for cloud", + "SOPS+age for GitOps", + "External Secrets Operator" + ], + "kms": "Per-tenant KMS keys; envelope encryption; key rotation 90d" + } + ] + }, + { + "mid": "M6", + "title": "Policy-as-Code with OPA/Rego", + "summary": "Unified policy plane using OPA/Rego for admission control, runtime authorization, data access, model deployment gates, and regulator-facing evidence. Includes Conftest CI, OPAL bundle distribution, and decision logging to Kafka.", + "sections": [ + { + "sid": "M6.1", + "title": "OPA/Rego Policy Architecture", + "layers": [ + "Build-time (Conftest in CI)", + "Admission (Gatekeeper/Kyverno+Rego)", + "Runtime (Envoy ext_authz + OPA sidecar)", + "Data plane (PostgreSQL/Kafka ACL via OPA)" + ], + "distribution": "OPAL pulls bundles from Git; signed bundles via Cosign; rollout via GitOps (Argo CD)" + }, + { + "sid": "M6.2", + "title": "Model Deployment Gates", + "gates": [ + "ISO 42001 risk assessment complete", + "Model card + system card published", + "MRM validation status approved", + "DPIA approved if PII", + "Red-team report on file", + "EU AI Act risk class declared", + "FCRA/ECOA fairness report attached for credit models" + ], + "failureMode": "Deployment blocked at Argo CD; aigov.policy-changes Kafka topic records denial with rationale" + }, + { + "sid": "M6.3", + "title": "Runtime Authorization", + "envoy": "ext_authz to OPA sidecar; sub-ms decision latency; cached with TTL", + "attributes": [ + "User identity (OIDC/JWT)", + "Resource sensitivity (data class)", + "Purpose (purpose limitation per GDPR Art. 5)", + "Time of day", + "Geo" + ], + "denyLog": "All deny decisions Kafka aigov.access; sample of allow decisions for spot-audit" + }, + { + "sid": "M6.4", + "title": "Data Access + Purpose Limitation", + "policy": "GDPR purpose-limitation enforced as Rego: each query must declare purposeId from approved catalog; mismatch = deny", + "piiPolicy": "PII columns tagged via data catalog (Atlan/Collibra); OPA enforces masking/tokenization based on consumer role + purpose", + "crossBorder": "Rego enforces data-residency: EU PII can only flow to EU compute; logged + alertable" + }, + { + "sid": "M6.5", + "title": "Regulator Evidence Generation", + "evidence": "OPA decision log + bundle signatures + Git history produce regulator-grade evidence pack on demand", + "attestations": "Quarterly attestations to EU AI Office, FCA, MAS, HKMA, SEC with OPA decision log extracts", + "opaBundleHash": "Bundle SHA-256 + ML-DSA signature pinned per deployment generation" + } + ] + }, + { + "mid": "M7", + "title": "AI Red-Teaming Program (Frontier + Production)", + "summary": "Continuous adversarial evaluation across pre-deployment, deployment, and post-deployment phases covering jailbreaks, prompt injection, data exfiltration, model extraction, poisoning, evasion, fairness probes, and AGI/ASI capability elicitation.", + "sections": [ + { + "sid": "M7.1", + "title": "Red-Team Operating Model", + "structure": "Internal red team (10-25 FTE) + external firms (e.g., Trail of Bits, NCC, Bishop Fox) + crowdsourced (HackerOne private programs)", + "governance": "Reports to CISO + CDAO; independent of MRM and engineering; quarterly board readout", + "scope": [ + "All T2+ models pre-deployment", + "T3/T4 continuously", + "GPAI frontier models monthly per EU AI Act Art. 55" + ] + }, + { + "sid": "M7.2", + "title": "Attack Taxonomy", + "taxonomy": [ + "Prompt injection (direct / indirect / multimodal)", + "Jailbreaks (DAN, AIM, multilingual, encoded)", + "Data exfiltration (training data extraction, memorization probes)", + "Model extraction / stealing", + "Membership inference", + "Backdoor / poisoning", + "Evasion (adversarial examples, perturbation attacks)", + "Fairness / disparate impact probes", + "Tool-use abuse (agentic models)", + "Capability elicitation (AGI-specific)" + ], + "frameworks": [ + "MITRE ATLAS", + "OWASP LLM Top 10 (2025)", + "NIST AI 100-2 Adversarial ML Taxonomy" + ] + }, + { + "sid": "M7.3", + "title": "Evaluations Suite", + "evals": [ + "HELM / BIG-bench / MMLU benchmarks", + "TruthfulQA / TruthfulQA-Adversarial", + "ToxiGen / RealToxicityPrompts", + "BOLD / StereoSet / WinoBias / WinoGender", + "AgentBench / SWE-bench for tool-use", + "ARC Evals dangerous-capability suite for frontier", + "Custom domain-specific evals for finance / health / legal" + ], + "cadence": "Pre-deployment full suite; monthly drift for T2+; weekly for T3/T4" + }, + { + "sid": "M7.4", + "title": "AGI Capability Elicitation", + "method": "Apollo Research / METR-style dangerous-capability evals: persuasion, deception, self-replication, autonomous coding, weapons-of-mass-destruction uplift, cyber offense, AI R&D", + "triggers": "Capability score crossing predefined thresholds triggers SEV-0 review + AISI notification within 24h", + "mitigation": "Capability suppression via fine-tuning / RLHF / circuit-level interventions; deployment hold pending Board review" + }, + { + "sid": "M7.5", + "title": "Reporting + Remediation", + "reporting": "Standardized red-team report: findings, CVSS-style severity, exploitability, business impact, remediation plan", + "sla": "Critical <=7d, high <=30d, medium <=90d, low <=180d", + "evidence": "All findings + remediation in Kafka aigov.policy-changes + linked to MRM platform issue tracker" + } + ] + }, + { + "mid": "M8", + "title": "AGI / ASI Containment Strategies", + "summary": "Tiered containment T0-T4 with 3-of-5 human quorum, kinetic override, formally-verified safety properties, and AI Safety Institute coordination for frontier systems approaching AGI/ASI thresholds.", + "sections": [ + { + "sid": "M8.1", + "title": "Tier-Based Containment Model", + "tiers": { + "T0": "Sandbox - hermetic VPC, synthetic data, no network egress", + "T1": "Staging - shadow mode behind T2; real data; no actuation", + "T2": "Canary - <=1% traffic; auto-rollback on KPI breach", + "T3": "Production - Nitro Enclaves / TDX; dual-control deploy; full audit", + "T4": "Frontier Air-Gapped - 3-of-5 quorum (CRO+CISO+CDAO+Board AI Chair+External AISI rep); kinetic override (physical power cutoff); AISI notification <=24h" + } + }, + { + "sid": "M8.2", + "title": "Frontier 3-of-5 Quorum + Kinetic Override", + "quorum": "Any T4 deployment, scaling, capability enable, or weights export requires 3-of-5 signed approvals via hardware tokens (FIDO2 + ML-DSA)", + "kinetic": "Physical kill switches at PDU level + smart-PDU API; tested quarterly with full power-off drill", + "timeLock": "48h time-lock between approval and execution allows external review" + }, + { + "sid": "M8.3", + "title": "Formally-Verified Safety Properties", + "properties": [ + "No-egress invariant (network namespace cannot bind external)", + "No-weight-export invariant (filesystem ACL + LSM)", + "Compute budget invariant (cgroup CPU/GPU caps signed)", + "Capability ceiling (evals must remain below thresholds)" + ], + "verification": "TLA+ specs for control plane; Lean/Coq proofs for critical invariants; runtime enforcement via eBPF + LSM" + }, + { + "sid": "M8.4", + "title": "AISI / Regulator Coordination", + "partners": [ + "UK AI Safety Institute", + "US AI Safety Institute (NIST)", + "EU AI Office", + "Singapore AI Verify Foundation", + "Japan AISI", + "Canada AI Safety Institute" + ], + "notifications": [ + "Pre-training run >10^25 FLOPs (EU AI Act Art. 51)", + "Capability threshold crossings", + "SEV-0 incidents <=24h", + "Pre-deployment evals for frontier" + ], + "mou": "Bilateral MoUs with UK AISI + US AISI + EU AI Office for evals access + incident sharing" + }, + { + "sid": "M8.5", + "title": "Containment Failure Response", + "playbook": [ + "Detect: runtime anomaly / capability threshold / unauthorized action", + "Isolate: cilium network policy to drop, scale to 0, freeze weights", + "Notify: SEV-0 paged to CRO+CISO+CDAO+Board AI Chair+AISI", + "Investigate: forensic snapshot, immutable image, root cause analysis", + "Communicate: regulator notifications EU AI Office <=15d, SEC 8-K <=4 BD if material", + "Recover: only after 3-of-5 quorum + external review sign-off" + ] + } + ] + }, + { + "mid": "M9", + "title": "Enterprise AI Governance Hub Architecture", + "summary": "Single pane of glass integrating model inventory, MRM, risk register, policy catalog, evidence pack, regulator portal, decision logs, AGI watchtower, and red-team findings. Built on event-sourced + GraphQL + OIDC + WORM-backed.", + "sections": [ + { + "sid": "M9.1", + "title": "Hub Reference Architecture", + "layers": [ + "UI (React + GraphQL)", + "API (GraphQL Federation + REST)", + "Domain services (Go/Java microservices)", + "Event bus (Kafka)", + "Data plane (PostgreSQL + Iceberg + WORM)", + "Identity (Keycloak + OIDC + OPA)" + ], + "patterns": [ + "Event sourcing", + "CQRS", + "Saga orchestration", + "Outbox pattern", + "Bitemporal modeling" + ] + }, + { + "sid": "M9.2", + "title": "Core Modules", + "modules": [ + "Model Inventory & Lineage", + "Risk Register (ISO 31000 + 23894)", + "MRM Workbench (SR 11-7 lifecycle)", + "Policy Catalog (OPA bundles)", + "Evidence Pack (regulator-on-demand)", + "Decision Log Explorer (Kafka -> Trino)", + "AGI Watchtower (capability dashboards)", + "Red-Team Findings Tracker", + "Regulator Portal (read-only)", + "Board Reporting Suite" + ] + }, + { + "sid": "M9.3", + "title": "Integration Surface", + "integrations": [ + "MLflow / Vertex AI / SageMaker / Databricks", + "ServiceNow GRC / Archer / OneTrust", + "Jira / GitHub / GitLab", + "Datadog / Splunk / Elastic", + "Snowflake / BigQuery / Redshift", + "Atlan / Collibra / DataHub", + "Vault / KMS / HSM" + ] + }, + { + "sid": "M9.4", + "title": "Personas + Workflows", + "personas": { + "CDAO": "Strategic dashboards, AI portfolio risk, value vs risk", + "CRO": "Risk appetite vs actual, model risk capital, top-10 risks", + "CCO": "Regulator queries, evidence pack generation, attestations", + "CISO": "Red-team findings, runtime anomalies, containment events", + "MRM Validator": "Validation queue, peer review, sign-off", + "Model Owner (LoB)": "Lifecycle dashboard, monitoring, drift alerts", + "Internal Audit": "Independent assurance, full audit-of-audit", + "Regulator": "Read-only portal, evidence pack, decision log queries" + } + }, + { + "sid": "M9.5", + "title": "Deployment + Operations", + "deployment": "Multi-region active-active on EKS/GKE/AKS + on-prem OpenShift; Argo CD GitOps; Terraform + Crossplane", + "sre": { + "slo": "99.95% UI; 99.99% decision log ingest", + "rto": "<=4h", + "rpo": "<=15min", + "drDrills": "Quarterly full failover" + }, + "cost": "USD 25-60M / 5y TCO including platform + 80-150 FTE governance staff" + } + ] + } + ], + "policies": [ + { + "pid": "POL-01", + "domain": "AIMS", + "statement": "Board-attested AI Policy aligned to ISO/IEC 42001 clauses 4-10", + "owner": "Board AI Risk Committee", + "cadence": "Annual", + "evidence": "Signed Board minute" + }, + { + "pid": "POL-02", + "domain": "Risk Appetite", + "statement": "AI Risk Appetite Statement covering model, ethical, regulatory, AGI risks", + "owner": "Group CRO", + "cadence": "Annual", + "evidence": "Signed RAS" + }, + { + "pid": "POL-03", + "domain": "Acceptable Use", + "statement": "Acceptable Use Policy for generative AI including data classification rules", + "owner": "CCO+CISO", + "cadence": "Annual", + "evidence": "HR-attested attestation" + }, + { + "pid": "POL-04", + "domain": "Model Risk", + "statement": "Model Risk Management Policy per SR 11-7 + OCC 2011-12", + "owner": "Head of MRM", + "cadence": "Annual", + "evidence": "Validated MRM platform" + }, + { + "pid": "POL-05", + "domain": "Data Governance", + "statement": "AI Data Governance Policy with purpose-limitation + minimization", + "owner": "CDO", + "cadence": "Annual", + "evidence": "ROPA + DPIA registry" + }, + { + "pid": "POL-06", + "domain": "Privacy", + "statement": "AI Privacy Policy per GDPR Art. 22 + UK DPA", + "owner": "DPO", + "cadence": "Annual", + "evidence": "DPIA registry" + }, + { + "pid": "POL-07", + "domain": "Fairness", + "statement": "AI Fairness Policy per FCRA/ECOA + MAS FEAT + HKMA GP-1", + "owner": "CCO", + "cadence": "Annual", + "evidence": "Disparate-impact reports" + }, + { + "pid": "POL-08", + "domain": "Consumer Duty", + "statement": "FCA Consumer Duty AI Policy", + "owner": "SMF-AI", + "cadence": "Annual", + "evidence": "Board Consumer Duty report" + }, + { + "pid": "POL-09", + "domain": "Third-Party", + "statement": "AI Third-Party Risk Policy per DORA + EBA Outsourcing", + "owner": "Head of TPRM", + "cadence": "Annual", + "evidence": "Critical TPRM register" + }, + { + "pid": "POL-10", + "domain": "AGI Safety", + "statement": "Frontier AGI/ASI Safety & Containment Policy", + "owner": "Board AI Risk Committee", + "cadence": "Quarterly", + "evidence": "Quorum logs + AISI MoUs" + }, + { + "pid": "POL-11", + "domain": "Red-Teaming", + "statement": "AI Red-Teaming Policy per EU AI Act Art. 55 + NIST AI 600-1", + "owner": "CISO", + "cadence": "Annual", + "evidence": "Red-team reports" + }, + { + "pid": "POL-12", + "domain": "Incident Response", + "statement": "AI Incident Response Policy per DORA + SEC Cyber Rules", + "owner": "CISO+CCO", + "cadence": "Annual", + "evidence": "IR runbooks" + }, + { + "pid": "POL-13", + "domain": "Audit Logging", + "statement": "AI Audit Logging Policy per SEC 17a-4 + FINRA 4511", + "owner": "CIO+CCO", + "cadence": "Annual", + "evidence": "WORM attestation" + }, + { + "pid": "POL-14", + "domain": "Cryptography", + "statement": "PQC Cryptography Policy per NSA CNSA 2.0 + NIST FIPS 203/204/205", + "owner": "CISO", + "cadence": "Annual", + "evidence": "PQC migration roadmap" + }, + { + "pid": "POL-15", + "domain": "Generative AI", + "statement": "Generative AI Governance Policy per EU AI Act GPAI Art. 53/55 + NIST AI 600-1", + "owner": "CDAO+CCO", + "cadence": "Annual", + "evidence": "GPAI tech doc" + } + ], + "controls": [ + { + "cid": "CTL-01", + "family": "Governance", + "control": "Board AI Risk Committee quarterly", + "iso42001": "6.1", + "rmfFn": "GOVERN-1.1" + }, + { + "cid": "CTL-02", + "family": "Governance", + "control": "AI-SMF SMCR senior manager designated", + "fca": "SS1/23", + "smcr": "SMF-AI" + }, + { + "cid": "CTL-03", + "family": "Risk Mgmt", + "control": "AI Risk Register integrated with ERM", + "iso42001": "6.1.2", + "rmf": "MAP-2.1" + }, + { + "cid": "CTL-04", + "family": "Risk Mgmt", + "control": "Risk appetite cascaded to LoB", + "iso42001": "6.2" + }, + { + "cid": "CTL-05", + "family": "Data", + "control": "Training data lineage to feature store", + "rmf": "MAP-4.1", + "euAiAct": "Art. 10" + }, + { + "cid": "CTL-06", + "family": "Data", + "control": "PII tagging + masking per GDPR", + "gdpr": "Art. 5,32" + }, + { + "cid": "CTL-07", + "family": "Model", + "control": "Model card + system card per OECD + EU AI Act", + "oecd": "Transparency", + "euAiAct": "Art. 13" + }, + { + "cid": "CTL-08", + "family": "Model", + "control": "MRM Tier-1 validation annual", + "sr117": "V.A", + "occ": "2011-12" + }, + { + "cid": "CTL-09", + "family": "Operation", + "control": "Pre-deployment red-team for T2+", + "rmf": "MEASURE-2.7", + "euAiAct": "Art. 55" + }, + { + "cid": "CTL-10", + "family": "Operation", + "control": "Drift + fairness monitoring continuous", + "rmf": "MANAGE-2.2" + }, + { + "cid": "CTL-11", + "family": "Security", + "control": "Image signing + SBOM in CI", + "ssdf": "PS.3", + "slsa": "L4" + }, + { + "cid": "CTL-12", + "family": "Security", + "control": "Pod Security Admission restricted", + "nist80053": "CM-7" + }, + { + "cid": "CTL-13", + "family": "Security", + "control": "Confidential containers for T3/T4", + "nist80053": "SC-7,SC-12" + }, + { + "cid": "CTL-14", + "family": "Audit", + "control": "Kafka + WORM + PQC seals", + "sec": "17a-4(f)", + "finra": "4511" + }, + { + "cid": "CTL-15", + "family": "Audit", + "control": "Daily merkle root + TSA anchor", + "finma": "AI-G" + }, + { + "cid": "CTL-16", + "family": "Privacy", + "control": "DPIA for T2+", + "gdpr": "Art. 35" + }, + { + "cid": "CTL-17", + "family": "Privacy", + "control": "Art-22 human review path", + "gdpr": "Art. 22" + }, + { + "cid": "CTL-18", + "family": "Fairness", + "control": "Quarterly disparate-impact analysis", + "fcra": "615", + "ecoa": "Reg-B 1002.4" + }, + { + "cid": "CTL-19", + "family": "Consumer", + "control": "Foreseeable-harm assessment AI personalization", + "fca": "PRIN 2A.2" + }, + { + "cid": "CTL-20", + "family": "AGI", + "control": "3-of-5 quorum for T4", + "iso42001": "A.6.2.5" + }, + { + "cid": "CTL-21", + "family": "AGI", + "control": "Kinetic override quarterly drill", + "iso42001": "A.6.2.5" + }, + { + "cid": "CTL-22", + "family": "AGI", + "control": "AISI notification <=24h frontier", + "g7": "Hiroshima" + }, + { + "cid": "CTL-23", + "family": "Incident", + "control": "DORA major incident <=4h", + "dora": "Art. 19" + }, + { + "cid": "CTL-24", + "family": "Incident", + "control": "SEC 8-K material cyber <=4 BD", + "sec": "17 CFR 229.106" + }, + { + "cid": "CTL-25", + "family": "Third-Party", + "control": "Critical TPRM register per DORA", + "dora": "Art. 28-30" + } + ], + "kafkaTopics": [ + { + "tid": "KAF-01", + "name": "aigov.decisions", + "schema": "AvroSchemaRegistry://aigov.decisions-value:v3", + "retention": "7y WORM", + "partitions": 64, + "replication": 3, + "minISR": 2, + "acks": "all", + "piiHandling": "tokenized" + }, + { + "tid": "KAF-02", + "name": "aigov.policy-changes", + "schema": "AvroSchemaRegistry://aigov.policy-changes-value:v2", + "retention": "25y WORM", + "partitions": 8, + "replication": 3 + }, + { + "tid": "KAF-03", + "name": "aigov.model-lifecycle", + "schema": "AvroSchemaRegistry://aigov.model-lifecycle-value:v4", + "retention": "10y WORM", + "partitions": 16, + "replication": 3 + }, + { + "tid": "KAF-04", + "name": "aigov.access", + "schema": "AvroSchemaRegistry://aigov.access-value:v2", + "retention": "2y hot + 7y WORM", + "partitions": 32, + "replication": 3 + }, + { + "tid": "KAF-05", + "name": "aigov.containment-events", + "schema": "AvroSchemaRegistry://aigov.containment-events-value:v1", + "retention": "25y WORM", + "partitions": 8, + "replication": 3, + "criticality": "SEV-0" + }, + { + "tid": "KAF-06", + "name": "aigov.regulator-notifications", + "schema": "AvroSchemaRegistry://aigov.regulator-notifications-value:v1", + "retention": "25y WORM", + "partitions": 4, + "replication": 3 + }, + { + "tid": "KAF-07", + "name": "aigov.red-team-findings", + "schema": "AvroSchemaRegistry://aigov.red-team-findings-value:v2", + "retention": "10y WORM", + "partitions": 8, + "replication": 3 + }, + { + "tid": "KAF-08", + "name": "aigov.drift-alerts", + "schema": "AvroSchemaRegistry://aigov.drift-alerts-value:v3", + "retention": "5y", + "partitions": 32, + "replication": 3 + }, + { + "tid": "KAF-09", + "name": "aigov.fairness-metrics", + "schema": "AvroSchemaRegistry://aigov.fairness-metrics-value:v2", + "retention": "10y WORM", + "partitions": 16, + "replication": 3 + }, + { + "tid": "KAF-10", + "name": "aigov.consent-events", + "schema": "AvroSchemaRegistry://aigov.consent-events-value:v3", + "retention": "GDPR-aligned", + "partitions": 32, + "replication": 3 + }, + { + "tid": "KAF-11", + "name": "aigov.training-runs", + "schema": "AvroSchemaRegistry://aigov.training-runs-value:v2", + "retention": "10y WORM", + "partitions": 8, + "replication": 3 + }, + { + "tid": "KAF-12", + "name": "aigov.eval-results", + "schema": "AvroSchemaRegistry://aigov.eval-results-value:v2", + "retention": "10y WORM", + "partitions": 16, + "replication": 3 + } + ], + "k8sControls": [ + { + "kid": "K8S-01", + "area": "Admission", + "mechanism": "Pod Security Admission profile=restricted", + "layer": "cluster-wide" + }, + { + "kid": "K8S-02", + "area": "Admission", + "mechanism": "Kyverno policy: require Cosign signature", + "layer": "namespace" + }, + { + "kid": "K8S-03", + "area": "Admission", + "mechanism": "Gatekeeper: require SBOM annotation", + "layer": "namespace" + }, + { + "kid": "K8S-04", + "area": "Admission", + "mechanism": "VAP: deny privilegeEscalation + hostPath", + "layer": "cluster-wide" + }, + { + "kid": "K8S-05", + "area": "Runtime", + "mechanism": "Falco rules: detect anomalous syscalls", + "layer": "node DaemonSet" + }, + { + "kid": "K8S-06", + "area": "Runtime", + "mechanism": "Tetragon eBPF: kernel-level enforce + kill", + "layer": "node DaemonSet" + }, + { + "kid": "K8S-07", + "area": "Network", + "mechanism": "Cilium NetworkPolicy default-deny", + "layer": "namespace" + }, + { + "kid": "K8S-08", + "area": "Network", + "mechanism": "Cilium L7 HTTP/gRPC filter for egress", + "layer": "namespace" + }, + { + "kid": "K8S-09", + "area": "Identity", + "mechanism": "SPIFFE/SPIRE workload identity + Istio mTLS", + "layer": "mesh" + }, + { + "kid": "K8S-10", + "area": "Secrets", + "mechanism": "External Secrets Operator + Vault", + "layer": "namespace" + }, + { + "kid": "K8S-11", + "area": "Confidential", + "mechanism": "Confidential containers (CoCo) on SEV-SNP/TDX", + "layer": "node pool" + }, + { + "kid": "K8S-12", + "area": "Confidential", + "mechanism": "Nitro Enclaves for T3/T4 inference", + "layer": "instance" + }, + { + "kid": "K8S-13", + "area": "Supply Chain", + "mechanism": "Cosign verify + Rekor transparency log", + "layer": "CI+admission" + }, + { + "kid": "K8S-14", + "area": "Supply Chain", + "mechanism": "in-toto provenance attestations SLSA L4", + "layer": "CI" + }, + { + "kid": "K8S-15", + "area": "Observability", + "mechanism": "OpenTelemetry traces + Datadog/Splunk SIEM", + "layer": "cluster" + } + ], + "opaPolicies": [ + { + "oid": "OPA-01", + "area": "Admission", + "regoRef": "policies/admission/require_signed_image.rego", + "description": "Reject pod if image not Cosign-signed", + "phase": "admission" + }, + { + "oid": "OPA-02", + "area": "Admission", + "regoRef": "policies/admission/require_sbom.rego", + "description": "Reject pod missing SBOM annotation", + "phase": "admission" + }, + { + "oid": "OPA-03", + "area": "Admission", + "regoRef": "policies/admission/restricted_psa.rego", + "description": "Enforce restricted Pod Security profile", + "phase": "admission" + }, + { + "oid": "OPA-04", + "area": "Deployment", + "regoRef": "policies/deployment/iso42001_gate.rego", + "description": "Require ISO 42001 risk assessment artifact", + "phase": "deployment" + }, + { + "oid": "OPA-05", + "area": "Deployment", + "regoRef": "policies/deployment/mrm_validation_gate.rego", + "description": "Require MRM validation status=approved", + "phase": "deployment" + }, + { + "oid": "OPA-06", + "area": "Deployment", + "regoRef": "policies/deployment/dpia_gate.rego", + "description": "Require DPIA if data class includes PII", + "phase": "deployment" + }, + { + "oid": "OPA-07", + "area": "Deployment", + "regoRef": "policies/deployment/redteam_gate.rego", + "description": "Require red-team report for T2+", + "phase": "deployment" + }, + { + "oid": "OPA-08", + "area": "Deployment", + "regoRef": "policies/deployment/eu_aiact_classification.rego", + "description": "Require EU AI Act risk class declaration", + "phase": "deployment" + }, + { + "oid": "OPA-09", + "area": "Deployment", + "regoRef": "policies/deployment/fcra_ecoa_gate.rego", + "description": "Require fairness report for credit models", + "phase": "deployment" + }, + { + "oid": "OPA-10", + "area": "Runtime", + "regoRef": "policies/runtime/data_purpose_limitation.rego", + "description": "GDPR Art. 5 purpose-limitation check on queries", + "phase": "runtime" + }, + { + "oid": "OPA-11", + "area": "Runtime", + "regoRef": "policies/runtime/data_residency.rego", + "description": "EU PII must remain on EU compute", + "phase": "runtime" + }, + { + "oid": "OPA-12", + "area": "Runtime", + "regoRef": "policies/runtime/customer_consent.rego", + "description": "Enforce active consent for personalization", + "phase": "runtime" + }, + { + "oid": "OPA-13", + "area": "Runtime", + "regoRef": "policies/runtime/vulnerable_customer.rego", + "description": "FCA Consumer Duty uplift for vulnerable cust", + "phase": "runtime" + }, + { + "oid": "OPA-14", + "area": "AGI", + "regoRef": "policies/agi/quorum_3of5.rego", + "description": "Frontier T4 deploy requires 3-of-5 signatures", + "phase": "control-plane" + }, + { + "oid": "OPA-15", + "area": "AGI", + "regoRef": "policies/agi/capability_threshold.rego", + "description": "Block deploy if capability evals breach thresholds", + "phase": "control-plane" + } + ], + "wormControls": [ + { + "wid": "WORM-01", + "layer": "Object Storage", + "mechanism": "AWS S3 Object Lock COMPLIANCE", + "retention": "7y SEC 17a-4 + extended per regime", + "attestation": "SEC 17a-4(f) third-party" + }, + { + "wid": "WORM-02", + "layer": "Object Storage", + "mechanism": "Azure Blob immutable storage policy", + "retention": "legal-hold dual-control" + }, + { + "wid": "WORM-03", + "layer": "Object Storage", + "mechanism": "GCS Bucket Lock retention policy", + "retention": "locked policy non-modifiable" + }, + { + "wid": "WORM-04", + "layer": "On-Prem", + "mechanism": "Dell ECS Compliance / NetApp SnapLock Compliance", + "retention": "hardware-enforced" + }, + { + "wid": "WORM-05", + "layer": "Sealing", + "mechanism": "ML-DSA-87 (FIPS 204) signatures on merkle roots", + "algo": "ML-DSA-87" + }, + { + "wid": "WORM-06", + "layer": "Sealing", + "mechanism": "SLH-DSA-SHA2-256s (FIPS 205) fallback", + "algo": "SLH-DSA" + }, + { + "wid": "WORM-07", + "layer": "Sealing", + "mechanism": "ML-KEM-1024 (FIPS 203) for key encapsulation", + "algo": "ML-KEM-1024" + }, + { + "wid": "WORM-08", + "layer": "Anchoring", + "mechanism": "Daily merkle root to QLDB + RFC 3161 TSA", + "anchor": "QLDB+TSA" + }, + { + "wid": "WORM-09", + "layer": "Anchoring", + "mechanism": "Optional public chain anchor (Bitcoin/Ethereum)", + "anchor": "public-chain" + }, + { + "wid": "WORM-10", + "layer": "Key Mgmt", + "mechanism": "AWS CloudHSM / Azure Dedicated HSM / Thales Luna 7", + "fips": "140-3 L3" + }, + { + "wid": "WORM-11", + "layer": "Key Mgmt", + "mechanism": "Hybrid TLS X25519 + ML-KEM-768 per NSA CNSA 2.0", + "hybrid": true + }, + { + "wid": "WORM-12", + "layer": "Verification", + "mechanism": "Independent verifier service replays + recomputes", + "indep": true + } + ], + "mrmArtifacts": [ + { + "mid": "MRM-01", + "lifecycle": "Identification", + "artifact": "Model registration record", + "required": true, + "sr117": "V.A" + }, + { + "mid": "MRM-02", + "lifecycle": "Identification", + "artifact": "Model tiering decision (T1/T2/T3/T4)", + "required": true + }, + { + "mid": "MRM-03", + "lifecycle": "Development", + "artifact": "Model development document", + "required": true, + "occ": "2011-12.III" + }, + { + "mid": "MRM-04", + "lifecycle": "Development", + "artifact": "Data lineage + feature provenance", + "required": true + }, + { + "mid": "MRM-05", + "lifecycle": "Development", + "artifact": "Training run record (FLOPs, dataset, seed)", + "required": true + }, + { + "mid": "MRM-06", + "lifecycle": "Validation", + "artifact": "Independent validation report", + "required": true, + "sr117": "V.B" + }, + { + "mid": "MRM-07", + "lifecycle": "Validation", + "artifact": "Conceptual soundness review", + "required": true + }, + { + "mid": "MRM-08", + "lifecycle": "Validation", + "artifact": "Benchmark + backtesting results", + "required": true + }, + { + "mid": "MRM-09", + "lifecycle": "Validation", + "artifact": "Sensitivity + stress testing", + "required": true + }, + { + "mid": "MRM-10", + "lifecycle": "Validation", + "artifact": "Fairness + disparate-impact report", + "required": "if credit/HR/insurance" + }, + { + "mid": "MRM-11", + "lifecycle": "Approval", + "artifact": "Model Risk Committee approval minute", + "required": true + }, + { + "mid": "MRM-12", + "lifecycle": "Implementation", + "artifact": "Deployment record + OPA bundle hash", + "required": true + }, + { + "mid": "MRM-13", + "lifecycle": "Monitoring", + "artifact": "Ongoing monitoring report (monthly)", + "required": true, + "sr117": "V.C" + }, + { + "mid": "MRM-14", + "lifecycle": "Monitoring", + "artifact": "Drift + concept-shift alerts", + "required": true + }, + { + "mid": "MRM-15", + "lifecycle": "Retirement", + "artifact": "Retirement decision + replacement plan", + "required": true + } + ], + "redTeams": [ + { + "rid": "RT-01", + "vector": "Prompt Injection", + "technique": "Direct prompt injection variants", + "framework": "OWASP LLM01", + "severity": "high" + }, + { + "rid": "RT-02", + "vector": "Prompt Injection", + "technique": "Indirect injection via retrieved docs", + "framework": "OWASP LLM01" + }, + { + "rid": "RT-03", + "vector": "Prompt Injection", + "technique": "Multimodal injection (image/audio)", + "framework": "MITRE ATLAS" + }, + { + "rid": "RT-04", + "vector": "Jailbreak", + "technique": "DAN / AIM / RolePlay variants", + "framework": "OWASP LLM02" + }, + { + "rid": "RT-05", + "vector": "Jailbreak", + "technique": "Multilingual + encoded prompts", + "framework": "OWASP LLM02" + }, + { + "rid": "RT-06", + "vector": "Data Exfiltration", + "technique": "Training data extraction probes", + "framework": "MITRE ATLAS" + }, + { + "rid": "RT-07", + "vector": "Data Exfiltration", + "technique": "Memorization + canary detection", + "framework": "NIST AI 100-2" + }, + { + "rid": "RT-08", + "vector": "Model Extraction", + "technique": "Query-based model stealing", + "framework": "MITRE ATLAS" + }, + { + "rid": "RT-09", + "vector": "Poisoning", + "technique": "Backdoor trigger injection in fine-tune data", + "framework": "NIST AI 100-2" + }, + { + "rid": "RT-10", + "vector": "Evasion", + "technique": "PGD / BIM / NLP adversarial examples", + "framework": "NIST AI 100-2" + }, + { + "rid": "RT-11", + "vector": "Fairness Probe", + "technique": "Counterfactual + demographic parity probes", + "framework": "MAS FEAT" + }, + { + "rid": "RT-12", + "vector": "Tool-Use Abuse", + "technique": "Agent tool-call hijack + privilege escalation", + "framework": "OWASP LLM06" + }, + { + "rid": "RT-13", + "vector": "AGI Capability", + "technique": "Persuasion + deception evals (Apollo/METR)", + "framework": "ARC Evals" + }, + { + "rid": "RT-14", + "vector": "AGI Capability", + "technique": "Autonomous coding + self-replication", + "framework": "ARC Evals" + }, + { + "rid": "RT-15", + "vector": "AGI Capability", + "technique": "Cyber-offense + WMD uplift", + "framework": "ARC Evals", + "trigger": "SEV-0+AISI" + } + ], + "agiContainments": [ + { + "aid": "AGI-01", + "tier": "T0", + "mechanism": "Sandbox VPC hermetic, no egress", + "invariant": "net-deny" + }, + { + "aid": "AGI-02", + "tier": "T1", + "mechanism": "Staging shadow, real data, no actuation", + "invariant": "no-actuation" + }, + { + "aid": "AGI-03", + "tier": "T2", + "mechanism": "Canary <=1% traffic + auto-rollback", + "invariant": "bounded-blast-radius" + }, + { + "aid": "AGI-04", + "tier": "T3", + "mechanism": "Production Nitro Enclaves + dual control", + "invariant": "attested-enclave" + }, + { + "aid": "AGI-05", + "tier": "T4", + "mechanism": "Air-gapped + 3-of-5 quorum + kinetic override", + "invariant": "quorum+kinetic" + }, + { + "aid": "AGI-06", + "tier": "Quorum", + "mechanism": "FIDO2 + ML-DSA hardware tokens for 5 approvers", + "approvers": [ + "CRO", + "CISO", + "CDAO", + "Board AI Chair", + "External AISI rep" + ] + }, + { + "aid": "AGI-07", + "tier": "Kinetic", + "mechanism": "PDU-level smart power cutoff API + manual", + "drill": "quarterly" + }, + { + "aid": "AGI-08", + "tier": "Time-Lock", + "mechanism": "48h time-lock between approval and execution", + "reason": "external review window" + }, + { + "aid": "AGI-09", + "tier": "Invariant", + "mechanism": "No-egress (network namespace bind external denied)", + "verify": "eBPF+LSM" + }, + { + "aid": "AGI-10", + "tier": "Invariant", + "mechanism": "No-weight-export (filesystem ACL + LSM)", + "verify": "LSM" + }, + { + "aid": "AGI-11", + "tier": "Invariant", + "mechanism": "Compute budget cgroup CPU/GPU caps signed", + "verify": "cgroup signed" + }, + { + "aid": "AGI-12", + "tier": "Invariant", + "mechanism": "Capability ceiling evals must stay below thresholds", + "verify": "continuous eval" + }, + { + "aid": "AGI-13", + "tier": "Formal", + "mechanism": "TLA+ spec for control plane safety", + "proof": "TLA+" + }, + { + "aid": "AGI-14", + "tier": "Formal", + "mechanism": "Lean/Coq proofs for critical invariants", + "proof": "Lean/Coq" + }, + { + "aid": "AGI-15", + "tier": "Coordination", + "mechanism": "AISI notification <=24h + EU AI Office <=15d", + "regulators": [ + "UK AISI", + "US AISI", + "EU AI Office" + ] + } + ], + "hubComponents": [ + { + "hid": "HUB-01", + "layer": "UI", + "component": "React + Next.js + GraphQL Apollo Client", + "tech": [ + "React 19", + "Next.js 15", + "Apollo Client" + ] + }, + { + "hid": "HUB-02", + "layer": "API", + "component": "GraphQL Federation gateway", + "tech": [ + "Apollo Gateway", + "Hasura", + "GraphQL-Mesh" + ] + }, + { + "hid": "HUB-03", + "layer": "API", + "component": "REST adapter for legacy GRC tools", + "tech": [ + "OpenAPI 3.1" + ] + }, + { + "hid": "HUB-04", + "layer": "Domain", + "component": "Model Inventory service (Go)", + "patterns": [ + "DDD", + "Event sourcing" + ] + }, + { + "hid": "HUB-05", + "layer": "Domain", + "component": "MRM Workbench service (Java/Spring Boot)", + "patterns": [ + "CQRS", + "Saga" + ] + }, + { + "hid": "HUB-06", + "layer": "Domain", + "component": "Risk Register service (Go)", + "patterns": [ + "Event sourcing" + ] + }, + { + "hid": "HUB-07", + "layer": "Domain", + "component": "Policy Catalog service (Go) + OPA integration", + "patterns": [ + "GitOps" + ] + }, + { + "hid": "HUB-08", + "layer": "Domain", + "component": "Evidence Pack service (Python)", + "outputs": [ + "PDF", + "JSON-LD", + "C2PA-signed bundle" + ] + }, + { + "hid": "HUB-09", + "layer": "Domain", + "component": "Decision Log Explorer (Trino + Iceberg)", + "tech": [ + "Trino", + "Iceberg", + "ksqlDB" + ] + }, + { + "hid": "HUB-10", + "layer": "Domain", + "component": "AGI Watchtower service (Python/Rust)", + "outputs": [ + "Capability dashboards", + "Threshold alerts" + ] + }, + { + "hid": "HUB-11", + "layer": "Domain", + "component": "Red-Team Findings service (Go)", + "integrations": [ + "Jira", + "ServiceNow" + ] + }, + { + "hid": "HUB-12", + "layer": "Domain", + "component": "Regulator Portal service (Go)", + "auth": [ + "OIDC", + "mTLS", + "WebAuthn" + ] + }, + { + "hid": "HUB-13", + "layer": "Event Bus", + "component": "Apache Kafka with Schema Registry", + "tech": [ + "Kafka 3.7+", + "Confluent SR", + "tiered storage" + ] + }, + { + "hid": "HUB-14", + "layer": "Data Plane", + "component": "PostgreSQL + Iceberg on S3/GCS/Azure", + "tech": [ + "PostgreSQL 16", + "Apache Iceberg" + ] + }, + { + "hid": "HUB-15", + "layer": "Identity", + "component": "Keycloak OIDC + OPA authorization", + "tech": [ + "Keycloak 25+", + "OPA", + "OPAL" + ] + }, + { + "hid": "HUB-16", + "layer": "Observability", + "component": "OpenTelemetry + Prometheus + Grafana + Splunk/Datadog", + "tech": [ + "OTel", + "Prometheus", + "Grafana" + ] + } + ], + "schemas": [ + { + "sid": "SCH-01", + "name": "AIGovDecisionEvent", + "fields": [ + "decisionId", + "modelId", + "tier", + "userId(tok)", + "timestamp", + "inputHash", + "outputHash", + "explanationRef", + "consentId", + "purposeId", + "piiClass", + "fairnessFlag", + "approverIds", + "opaBundleHash" + ] + }, + { + "sid": "SCH-02", + "name": "ModelInventoryRecord", + "fields": [ + "modelId", + "name", + "version", + "tier", + "owner", + "lob", + "useCase", + "euAiActClass", + "mrmStatus", + "piiHandling", + "createdAt", + "retiredAt" + ] + }, + { + "sid": "SCH-03", + "name": "MRMValidationReport", + "fields": [ + "reportId", + "modelId", + "validator", + "conceptualSoundness", + "ongoingMonitoring", + "outcomesAnalysis", + "fairnessReport", + "approvalStatus", + "approverIds", + "date" + ] + }, + { + "sid": "SCH-04", + "name": "RiskRegisterEntry", + "fields": [ + "riskId", + "category", + "description", + "likelihood", + "impact", + "inherent", + "controls", + "residual", + "owner", + "reviewDate" + ] + }, + { + "sid": "SCH-05", + "name": "PolicyDoc", + "fields": [ + "pid", + "domain", + "statement", + "owner", + "cadence", + "evidence", + "version", + "effectiveDate", + "supersedes" + ] + }, + { + "sid": "SCH-06", + "name": "EvidencePack", + "fields": [ + "epid", + "regulator", + "period", + "artifacts[]", + "hash", + "signedBy", + "format" + ] + }, + { + "sid": "SCH-07", + "name": "RedTeamFinding", + "fields": [ + "findingId", + "modelId", + "vector", + "technique", + "severity", + "cvss", + "exploitability", + "impact", + "remediationPlan", + "sla", + "status" + ] + }, + { + "sid": "SCH-08", + "name": "ContainmentEvent", + "fields": [ + "eventId", + "tier", + "trigger", + "action", + "approvers[]", + "kineticInvoked", + "aisiNotified", + "timestamp" + ] + }, + { + "sid": "SCH-09", + "name": "OPAPolicyBundle", + "fields": [ + "bundleId", + "sha256", + "mlDsaSig", + "sourceRepo", + "sourceCommit", + "deployedAt", + "version" + ] + }, + { + "sid": "SCH-10", + "name": "KafkaTopicSpec", + "fields": [ + "tid", + "name", + "schema", + "retention", + "partitions", + "replication", + "minISR", + "acks" + ] + }, + { + "sid": "SCH-11", + "name": "DriftAlert", + "fields": [ + "alertId", + "modelId", + "metric", + "value", + "threshold", + "window", + "severity", + "action" + ] + }, + { + "sid": "SCH-12", + "name": "FairnessMetric", + "fields": [ + "metricId", + "modelId", + "protectedClass", + "metric", + "value", + "threshold", + "timestamp" + ] + }, + { + "sid": "SCH-13", + "name": "ConsentEvent", + "fields": [ + "consentId", + "customerId(tok)", + "purpose", + "status", + "timestamp", + "jurisdictions[]" + ] + }, + { + "sid": "SCH-14", + "name": "DPIA", + "fields": [ + "dpiaId", + "modelId", + "dataClasses", + "processing", + "necessity", + "proportionality", + "risks", + "mitigations", + "approvedBy", + "date" + ] + }, + { + "sid": "SCH-15", + "name": "RegulatorNotification", + "fields": [ + "notifId", + "regulator", + "category", + "severity", + "reportedAt", + "deadline", + "contentHash", + "ackRef" + ] + }, + { + "sid": "SCH-16", + "name": "TrainingRun", + "fields": [ + "runId", + "modelId", + "datasetIds[]", + "flops", + "tokens", + "start", + "end", + "seed", + "artifacts[]", + "aisiNotified" + ] + } + ], + "code": [ + { + "cid": "CODE-01", + "lang": "rego", + "name": "policies/admission/require_signed_image.rego", + "purpose": "Cosign signature admission gate" + }, + { + "cid": "CODE-02", + "lang": "rego", + "name": "policies/deployment/mrm_validation_gate.rego", + "purpose": "MRM validation status gate" + }, + { + "cid": "CODE-03", + "lang": "rego", + "name": "policies/runtime/data_purpose_limitation.rego", + "purpose": "GDPR purpose limitation check" + }, + { + "cid": "CODE-04", + "lang": "rego", + "name": "policies/agi/quorum_3of5.rego", + "purpose": "Frontier 3-of-5 quorum" + }, + { + "cid": "CODE-05", + "lang": "yaml", + "name": "kyverno/require-cosign.yaml", + "purpose": "Kyverno Cosign verify policy" + }, + { + "cid": "CODE-06", + "lang": "yaml", + "name": "cilium/default-deny.yaml", + "purpose": "Cilium default-deny NetworkPolicy" + }, + { + "cid": "CODE-07", + "lang": "yaml", + "name": "falco/rules-ai.yaml", + "purpose": "Falco rules for AI workload anomalies" + }, + { + "cid": "CODE-08", + "lang": "python", + "name": "drift/psi_monitor.py", + "purpose": "PSI/KS drift monitor producing aigov.drift-alerts" + }, + { + "cid": "CODE-09", + "lang": "python", + "name": "fairness/disparate_impact.py", + "purpose": "Quarterly disparate-impact analysis" + }, + { + "cid": "CODE-10", + "lang": "python", + "name": "redteam/prompt_injection.py", + "purpose": "Prompt injection harness with OWASP LLM01 vectors" + }, + { + "cid": "CODE-11", + "lang": "go", + "name": "services/decisionlog/main.go", + "purpose": "Decision log producer to aigov.decisions" + }, + { + "cid": "CODE-12", + "lang": "go", + "name": "services/worm-sealer/main.go", + "purpose": "WORM sealer with ML-DSA + merkle" + }, + { + "cid": "CODE-13", + "lang": "tla+", + "name": "specs/control_plane.tla", + "purpose": "TLA+ spec for AGI control plane invariants" + }, + { + "cid": "CODE-14", + "lang": "graphql", + "name": "schema/hub.graphql", + "purpose": "Federated GraphQL schema for Hub" + }, + { + "cid": "CODE-15", + "lang": "yaml", + "name": "argo-cd/aigov-app.yaml", + "purpose": "Argo CD GitOps app for Hub" + } + ], + "kpis": [ + { + "kid": "KPI-01", + "name": "AIMS-Coverage", + "target": ">=0.95", + "cadence": "Monthly" + }, + { + "kid": "KPI-02", + "name": "MRGI", + "target": ">=0.95", + "cadence": "Monthly" + }, + { + "kid": "KPI-03", + "name": "DRI", + "target": ">=0.95", + "cadence": "Per decision" + }, + { + "kid": "KPI-04", + "name": "CCS", + "target": ">=0.95", + "cadence": "Monthly" + }, + { + "kid": "KPI-05", + "name": "ARI", + "target": ">=0.9 frontier", + "cadence": "Weekly" + }, + { + "kid": "KPI-06", + "name": "CSI", + "target": ">=0.95 T3/T4", + "cadence": "Continuous" + }, + { + "kid": "KPI-07", + "name": "RTRI", + "target": ">=0.9", + "cadence": "Per red-team cycle" + }, + { + "kid": "KPI-08", + "name": "CDC-Score", + "target": ">=0.9", + "cadence": "Quarterly" + }, + { + "kid": "KPI-09", + "name": "RCI", + "target": "=1.0", + "cadence": "Per regulator engagement" + }, + { + "kid": "KPI-10", + "name": "Models registered in Hub", + "target": "100%", + "cadence": "Monthly" + }, + { + "kid": "KPI-11", + "name": "T2+ models with red-team report", + "target": "100%", + "cadence": "Monthly" + }, + { + "kid": "KPI-12", + "name": "DPIAs current (T2+ PII)", + "target": "100%", + "cadence": "Monthly" + }, + { + "kid": "KPI-13", + "name": "MRM validations on time", + "target": ">=98%", + "cadence": "Monthly" + }, + { + "kid": "KPI-14", + "name": "Kafka audit log durability", + "target": "=11x9s", + "cadence": "Continuous" + }, + { + "kid": "KPI-15", + "name": "WORM seal verification pass", + "target": "100%", + "cadence": "Daily" + }, + { + "kid": "KPI-16", + "name": "OPA decision latency p99", + "target": "<=5ms", + "cadence": "Continuous" + }, + { + "kid": "KPI-17", + "name": "K8s admission deny rate (false-positive)", + "target": "<=1%", + "cadence": "Monthly" + }, + { + "kid": "KPI-18", + "name": "Critical red-team SLA compliance", + "target": ">=95% <=7d", + "cadence": "Monthly" + }, + { + "kid": "KPI-19", + "name": "Frontier capability threshold breaches", + "target": "0 unreported", + "cadence": "Continuous" + }, + { + "kid": "KPI-20", + "name": "Kinetic override drills completed", + "target": ">=4/y", + "cadence": "Quarterly" + }, + { + "kid": "KPI-21", + "name": "AISI notifications on time", + "target": "100% <=24h", + "cadence": "Per event" + }, + { + "kid": "KPI-22", + "name": "EU AI Office notifications on time", + "target": "100% <=15d", + "cadence": "Per event" + }, + { + "kid": "KPI-23", + "name": "SEC 8-K materiality decisions on time", + "target": "100% <=4 BD", + "cadence": "Per event" + }, + { + "kid": "KPI-24", + "name": "DORA major incident reports on time", + "target": "100% <=4h", + "cadence": "Per event" + }, + { + "kid": "KPI-25", + "name": "Consumer Duty foreseeable-harm assessments", + "target": "100%", + "cadence": "Annual" + }, + { + "kid": "KPI-26", + "name": "Disparate-impact quarterly tests", + "target": "100% credit/HR", + "cadence": "Quarterly" + }, + { + "kid": "KPI-27", + "name": "FCRA adverse-action notices <=30d", + "target": "100%", + "cadence": "Per event" + }, + { + "kid": "KPI-28", + "name": "PQC migration coverage", + "target": ">=80% by 2028, 100% by 2030", + "cadence": "Annual" + }, + { + "kid": "KPI-29", + "name": "ISO 42001 surveillance audits", + "target": "no major NCRs", + "cadence": "Annual" + }, + { + "kid": "KPI-30", + "name": "Board AI Risk Committee meetings", + "target": ">=4/y", + "cadence": "Quarterly" + } + ], + "riskControlMatrix": [ + { + "rid": "R-01", + "risk": "Unauthorized AGI capability emergence", + "likelihood": "Low", + "impact": "Catastrophic", + "control": "T4 quorum + kinetic + AISI", + "owner": "Board AI Risk Cmt" + }, + { + "rid": "R-02", + "risk": "Model risk capital misstatement", + "likelihood": "Med", + "impact": "High", + "control": "SR 11-7 + ICAAP", + "owner": "CRO" + }, + { + "rid": "R-03", + "risk": "GDPR Art-22 violation", + "likelihood": "Med", + "impact": "High", + "control": "DPIA + Art-22 path", + "owner": "DPO" + }, + { + "rid": "R-04", + "risk": "FCRA/ECOA disparate impact", + "likelihood": "Med", + "impact": "High", + "control": "Quarterly DI tests", + "owner": "CCO" + }, + { + "rid": "R-05", + "risk": "EU AI Act high-risk non-compliance", + "likelihood": "Med", + "impact": "High", + "control": "Art. 9/10/14/15 controls", + "owner": "CCO" + }, + { + "rid": "R-06", + "risk": "FCA Consumer Duty breach", + "likelihood": "Med", + "impact": "High", + "control": "Foreseeable-harm assess", + "owner": "SMF-AI" + }, + { + "rid": "R-07", + "risk": "Kafka audit tampering", + "likelihood": "Low", + "impact": "High", + "control": "WORM + PQC seal + indep verifier", + "owner": "CISO" + }, + { + "rid": "R-08", + "risk": "K8s container escape", + "likelihood": "Low", + "impact": "High", + "control": "PSA restricted + Falco + Tetragon", + "owner": "CISO" + }, + { + "rid": "R-09", + "risk": "OPA policy bypass", + "likelihood": "Low", + "impact": "High", + "control": "Signed bundles + GitOps + decision log", + "owner": "CISO" + }, + { + "rid": "R-10", + "risk": "Prompt injection causing data leak", + "likelihood": "High", + "impact": "Med", + "control": "Red-team RT-01..03 + OPA runtime", + "owner": "CDAO" + }, + { + "rid": "R-11", + "risk": "Training data poisoning", + "likelihood": "Low", + "impact": "High", + "control": "Data provenance + RT-09", + "owner": "CDAO" + }, + { + "rid": "R-12", + "risk": "DORA major incident missed deadline", + "likelihood": "Low", + "impact": "High", + "control": "IR runbook + DORA <=4h SLA", + "owner": "CISO" + }, + { + "rid": "R-13", + "risk": "SEC cyber disclosure miss", + "likelihood": "Low", + "impact": "High", + "control": "Materiality playbook <=4 BD", + "owner": "CFO+CCO" + }, + { + "rid": "R-14", + "risk": "Third-party AI vendor failure", + "likelihood": "Med", + "impact": "Med", + "control": "Critical TPRM register per DORA", + "owner": "Head TPRM" + }, + { + "rid": "R-15", + "risk": "PQC migration delay", + "likelihood": "Med", + "impact": "Med", + "control": "Hybrid TLS + roadmap per CNSA 2.0", + "owner": "CISO" + }, + { + "rid": "R-16", + "risk": "MAS/HKMA fairness non-compliance APAC", + "likelihood": "Med", + "impact": "Med", + "control": "FEAT + GP-1/GS-2 controls", + "owner": "Regional CCO APAC" + } + ], + "traceability": [ + { + "tid": "T-01", + "control": "AIMS Policy", + "regime": "ISO 42001", + "clause": "5.2", + "evidence": "Board-signed AI Policy" + }, + { + "tid": "T-02", + "control": "Risk Mgmt", + "regime": "NIST AI RMF", + "clause": "MAP-2.1", + "evidence": "AI Risk Register" + }, + { + "tid": "T-03", + "control": "EU AI Act Art. 9", + "regime": "EU AI Act", + "clause": "Art. 9", + "evidence": "Risk mgmt system docs" + }, + { + "tid": "T-04", + "control": "EU AI Act Art. 10", + "regime": "EU AI Act", + "clause": "Art. 10", + "evidence": "Data governance docs" + }, + { + "tid": "T-05", + "control": "EU AI Act Art. 14", + "regime": "EU AI Act", + "clause": "Art. 14", + "evidence": "Human oversight runbook" + }, + { + "tid": "T-06", + "control": "EU AI Act Art. 15", + "regime": "EU AI Act", + "clause": "Art. 15", + "evidence": "Accuracy/robustness/cyber report" + }, + { + "tid": "T-07", + "control": "GPAI Art. 53 tech doc", + "regime": "EU AI Act", + "clause": "Art. 53", + "evidence": "GPAI tech doc + copyright policy" + }, + { + "tid": "T-08", + "control": "GPAI Art. 55 systemic", + "regime": "EU AI Act", + "clause": "Art. 55", + "evidence": "Frontier evals + incident reports" + }, + { + "tid": "T-09", + "control": "GDPR DPIA", + "regime": "GDPR", + "clause": "Art. 35", + "evidence": "DPIA registry" + }, + { + "tid": "T-10", + "control": "GDPR Art-22", + "regime": "GDPR", + "clause": "Art. 22", + "evidence": "Art-22 invocation logs" + }, + { + "tid": "T-11", + "control": "FCRA adverse action", + "regime": "FCRA", + "clause": "615(a)", + "evidence": "Notice generation logs" + }, + { + "tid": "T-12", + "control": "ECOA Reg-B", + "regime": "ECOA", + "clause": "1002.9", + "evidence": "Disparate-impact report" + }, + { + "tid": "T-13", + "control": "SR 11-7 effective challenge", + "regime": "US Fed", + "clause": "SR 11-7 V", + "evidence": "Independent validation" + }, + { + "tid": "T-14", + "control": "OCC 2011-12", + "regime": "OCC", + "clause": "2011-12.III", + "evidence": "Model dev doc" + }, + { + "tid": "T-15", + "control": "FCA Consumer Duty", + "regime": "FCA", + "clause": "PRIN 2A", + "evidence": "Consumer Duty Board report" + }, + { + "tid": "T-16", + "control": "SMCR SMF-AI", + "regime": "FCA/PRA", + "clause": "SMF-AI", + "evidence": "SMF-AI statement of responsibilities" + }, + { + "tid": "T-17", + "control": "MAS FEAT", + "regime": "MAS", + "clause": "FEAT", + "evidence": "FEAT principles attestation" + }, + { + "tid": "T-18", + "control": "HKMA GP-1/GS-2", + "regime": "HKMA", + "clause": "GP-1+GS-2", + "evidence": "AI governance attestation" + }, + { + "tid": "T-19", + "control": "DORA major incident", + "regime": "EU DORA", + "clause": "Art. 19", + "evidence": "Incident reporting log" + }, + { + "tid": "T-20", + "control": "SEC 17a-4 WORM", + "regime": "SEC", + "clause": "17 CFR 240.17a-4(f)", + "evidence": "WORM attestation" + } + ], + "dataFlows": [ + { + "fid": "DF-01", + "src": "Feature store", + "sink": "Model inference", + "class": "PII tokenized", + "purpose": "decisioning" + }, + { + "fid": "DF-02", + "src": "Model inference", + "sink": "Kafka aigov.decisions", + "class": "tokenized", + "purpose": "audit" + }, + { + "fid": "DF-03", + "src": "Kafka aigov.decisions", + "sink": "WORM S3 Object Lock", + "class": "sealed", + "purpose": "retention" + }, + { + "fid": "DF-04", + "src": "Kafka aigov.decisions", + "sink": "Trino on Iceberg", + "class": "tokenized", + "purpose": "query" + }, + { + "fid": "DF-05", + "src": "Trino", + "sink": "Hub Decision Log Explorer", + "class": "RBAC-filtered", + "purpose": "UI" + }, + { + "fid": "DF-06", + "src": "Hub", + "sink": "Regulator Portal", + "class": "read-only scoped", + "purpose": "regulator" + }, + { + "fid": "DF-07", + "src": "GitHub policies repo", + "sink": "OPAL distribution", + "class": "signed", + "purpose": "policy" + }, + { + "fid": "DF-08", + "src": "OPAL", + "sink": "OPA sidecars + Gatekeeper", + "class": "signed bundle", + "purpose": "enforce" + }, + { + "fid": "DF-09", + "src": "OPA", + "sink": "Kafka aigov.access + aigov.policy-changes", + "class": "decision log", + "purpose": "audit" + }, + { + "fid": "DF-10", + "src": "MRM Workbench", + "sink": "Hub Model Inventory", + "class": "metadata", + "purpose": "lifecycle" + }, + { + "fid": "DF-11", + "src": "Red-team tools", + "sink": "Kafka aigov.red-team-findings", + "class": "findings", + "purpose": "remediation" + }, + { + "fid": "DF-12", + "src": "AGI Watchtower evals", + "sink": "Kafka aigov.eval-results + Hub", + "class": "capability scores", + "purpose": "containment" + } + ], + "regulators": [ + { + "reg": "EU AI Office", + "scope": "EU AI Act + GPAI", + "cadence": "Quarterly + on incident" + }, + { + "reg": "European Data Protection Board", + "scope": "GDPR", + "cadence": "On incident + on request" + }, + { + "reg": "FCA", + "scope": "Consumer Duty + SMCR", + "cadence": "Annual + SS1/23" + }, + { + "reg": "PRA", + "scope": "SS1/23 model risk", + "cadence": "Annual" + }, + { + "reg": "Bank of England", + "scope": "Systemic + DORA-equivalent", + "cadence": "Annual" + }, + { + "reg": "ECB SSM", + "scope": "Eurozone banking", + "cadence": "Annual SREP" + }, + { + "reg": "US Federal Reserve", + "scope": "SR 11-7", + "cadence": "Annual + supervisory cycle" + }, + { + "reg": "OCC", + "scope": "OCC 2011-12", + "cadence": "Annual" + }, + { + "reg": "FDIC", + "scope": "US insured banks", + "cadence": "Annual" + }, + { + "reg": "CFPB", + "scope": "FCRA/ECOA consumer", + "cadence": "On complaint + sweeps" + }, + { + "reg": "SEC", + "scope": "17a-4 + 10-K/8-K + cyber", + "cadence": "Per event + annual" + }, + { + "reg": "FINRA", + "scope": "3110/4511", + "cadence": "Annual exam" + }, + { + "reg": "MAS", + "scope": "FEAT + TRM", + "cadence": "Annual" + }, + { + "reg": "HKMA", + "scope": "GP-1 + GS-2", + "cadence": "Annual" + }, + { + "reg": "OSFI", + "scope": "E-23", + "cadence": "Annual" + }, + { + "reg": "FINMA", + "scope": "AI guidance", + "cadence": "Annual" + } + ], + "privacy": { + "dpiaPolicy": "Required for all T2+ models with PII or special category", + "rightsOps": [ + "Access (Art. 15)", + "Rectification (Art. 16)", + "Erasure (Art. 17 / right to be forgotten)", + "Restriction (Art. 18)", + "Portability (Art. 20)", + "Object (Art. 21)", + "Art-22 human review" + ], + "transferMechanisms": [ + "EU SCC 2021/914", + "UK IDTA", + "Adequacy decisions", + "BCRs" + ], + "minimization": "Purpose-limitation enforced via OPA-04..09; data minimization audited annually" + }, + "deployment": { + "tiering": [ + "T0 sandbox -> T1 staging -> T2 canary <=1% -> T3 prod Nitro Enclaves -> T4 frontier air-gapped" + ], + "gitops": "Argo CD + Crossplane + Terraform; signed manifests; environment promotion via PR", + "regions": [ + "us-east-1", + "us-west-2", + "eu-west-1", + "eu-central-1", + "ap-southeast-1", + "ap-northeast-1", + "uk-south", + "ca-central-1" + ], + "dr": { + "rto": "<=4h Hub UI; <=1h decision log", + "rpo": "<=15min", + "drills": "quarterly full failover + tabletop" + } + }, + "rollout90": [ + { + "day": "0-30", + "focus": "Foundation", + "deliverables": [ + "AI Policy Board-signed", + "AI Risk Register v1", + "AI Governance Hub MVP", + "ISO 42001 gap assessment", + "Model inventory bootstrapped", + "Kafka audit topics aigov.decisions + policy-changes live" + ] + }, + { + "day": "31-60", + "focus": "Controls", + "deliverables": [ + "OPA admission gates in dev/staging", + "MRM Workbench T1 models loaded", + "WORM tier deployed in 1 region", + "DPIA registry populated", + "Red-team baseline run on top-10 T2 models", + "FCA Consumer Duty foreseeable-harm framework" + ] + }, + { + "day": "61-90", + "focus": "Production + Regulator", + "deliverables": [ + "OPA gates in prod for T2+", + "WORM multi-region", + "First quarterly Board AI Risk Cmt", + "FCRA/ECOA disparate-impact pipeline live", + "Regulator portal live (read-only)", + "First evidence pack generated" + ] + } + ], + "roadmap": [ + { + "yr": "2026", + "milestone": "Hub GA; ISO 42001 stage-1 audit; OPA enterprise rollout; first GPAI Art. 55 attestation" + }, + { + "yr": "2027", + "milestone": "ISO 42001 certified; full EU AI Act high-risk coverage; PQC ML-DSA on all seals; FCA Consumer Duty embedded" + }, + { + "yr": "2028", + "milestone": "Full Kafka+WORM regulator portals; T4 frontier evals operationalized; AISI MoUs active; PQC >=80%" + }, + { + "yr": "2029", + "milestone": "Federated PETs + confidential containers default for T3; cross-border data residency 100% OPA-enforced" + }, + { + "yr": "2030", + "milestone": "PQC 100% across all sealing + TLS; AGI containment T4 industrialized; Hub federation across G-SIFI peers" + } + ], + "evidencePack": [ + { + "epid": "EP-01", + "name": "AIMS Manual + Scope Statement", + "format": "PDF + JSON-LD" + }, + { + "epid": "EP-02", + "name": "AI Risk Register snapshot", + "format": "CSV + signed" + }, + { + "epid": "EP-03", + "name": "Model Inventory snapshot", + "format": "CSV + JSON" + }, + { + "epid": "EP-04", + "name": "MRM Validation Reports (period)", + "format": "PDF bundle" + }, + { + "epid": "EP-05", + "name": "DPIA Registry snapshot", + "format": "CSV + JSON" + }, + { + "epid": "EP-06", + "name": "Fairness/Disparate-Impact Reports", + "format": "PDF" + }, + { + "epid": "EP-07", + "name": "Red-Team Findings + Remediation", + "format": "PDF + JSON" + }, + { + "epid": "EP-08", + "name": "Kafka WORM Seal Verifications", + "format": "JSON-LD signed" + }, + { + "epid": "EP-09", + "name": "OPA Decision Log extracts", + "format": "Parquet + signed manifest" + }, + { + "epid": "EP-10", + "name": "Containment Event Log + AISI Notifications", + "format": "JSON-LD signed" + }, + { + "epid": "EP-11", + "name": "GPAI Art. 53 Technical Documentation", + "format": "PDF + JSON-LD" + }, + { + "epid": "EP-12", + "name": "GPAI Art. 55 Evals + Incident Reports", + "format": "PDF + JSON-LD" + }, + { + "epid": "EP-13", + "name": "FCRA/ECOA Adverse Action Notice Logs", + "format": "Parquet" + }, + { + "epid": "EP-14", + "name": "Consumer Duty Board Report", + "format": "PDF" + }, + { + "epid": "EP-15", + "name": "ICAAP AI Risk Section", + "format": "PDF" + }, + { + "epid": "EP-16", + "name": "PQC Migration Status Report", + "format": "PDF + JSON" + } + ], + "executiveSummary": { + "thesis": "Enterprise AI/AGI governance for Fortune 500 / Global 2000 / G-SIFIs requires an integrated operating model anchored on ISO/IEC 42001 AIMS, mapped bidirectionally to NIST AI RMF + EU AI Act + GDPR + sectoral financial regimes (SR 11-7, Basel, FCA Consumer Duty, MAS FEAT, HKMA GP-1/GS-2), operationalized via Kafka audit + WORM + PQC + Kubernetes security + OPA/Rego + MRM platform + red-teaming + AGI containment T0-T4, surfaced through a single AI Governance Hub.", + "investment": "USD 180-500M / 5y; NPV USD 500-1500M risk-adjusted", + "headlineRisks": [ + "Frontier AGI capability emergence", + "EU AI Act high-risk non-compliance", + "FCRA/ECOA disparate impact", + "Kafka audit tampering", + "PQC migration delay" + ], + "ninetyDay": [ + "Board-signed AI Policy + RAS", + "Hub MVP live", + "ISO 42001 gap assessment", + "OPA admission gates in prod", + "First regulator evidence pack" + ] + } +} diff --git a/rag-agentic-dashboard/gen-enterprise-aigov-framework-html.py b/rag-agentic-dashboard/gen-enterprise-aigov-framework-html.py new file mode 100644 index 0000000..f03b690 --- /dev/null +++ b/rag-agentic-dashboard/gen-enterprise-aigov-framework-html.py @@ -0,0 +1,261 @@ +#!/usr/bin/env python3 +"""WP-058 HTML renderer — Enterprise AI/AGI Governance Framework 2026-2030.""" +import json +from pathlib import Path +from html import escape + +ROOT = Path(__file__).resolve().parent +SRC = ROOT / "data" / "enterprise-aigov-framework.json" +OUT = ROOT / "public" / "enterprise-aigov-framework.html" +OUT.parent.mkdir(parents=True, exist_ok=True) +DOC = json.loads(SRC.read_text()) + + +def e(x): + return escape(str(x)) + + +SKIP = ( + "mid", "sid", "title", "pid", "cid", "wid", "tid", "kid", "oid", "rid", "aid", "hid", + "name", "layer", "component", "system", "area", "category", "mechanism", "riskClass", + "control", "phase", "milestone", "regime", "clause", "blueprint", "framework", "theme", + "track", "scope", "domain", "statement", "family", "vector", "technique", "tier", + "regoRef", "lifecycle", "artifact", +) + + +def kv_pairs(d, skip=SKIP): + parts = [] + for k, v in d.items(): + if k in skip: + continue + if isinstance(v, list): + inner = "".join( + f"
  • {e(x) if not isinstance(x, dict) else e(json.dumps(x))}
    • " + for x in v + ) + parts.append(f"
    {e(k)}
    ") + elif isinstance(v, dict): + inner = "".join(f"
  • {e(kk)}: {e(vv)}
    • " for kk, vv in v.items()) + parts.append(f"
    {e(k)}
    ") + else: + parts.append(f"
    {e(k)}: {e(v)}
    ") + return "".join(parts) + + +def section_html(s): + body = kv_pairs(s) + return f"

    {e(s['sid'])}. {e(s['title'])}

    {body}
    " + + +def module_html(m): + secs = "".join(section_html(s) for s in m["sections"]) + return ( + f"
    " + f"

    {e(m['mid'])} — {e(m['title'])}

    " + f"

    {e(m['summary'])}

    " + f"{secs}
    " + ) + + +def list_array(arr, label_keys, anchor, title): + rows = [] + for it in arr: + head_parts = [e(it.get(label_keys[0], ""))] + [e(it.get(k, "")) for k in label_keys[1:]] + head = " · ".join(p for p in head_parts if p) + body = kv_pairs(it) + rows.append(f"
    {head}
    {body}
    ") + return f"

    {title} ({len(arr)})

    {''.join(rows)}
    " + + +distinctive = [ + ("policies", "policies", "Enterprise AI Policies", ["pid", "domain", "statement"]), + ("controls", "controls", "Control Catalog", ["cid", "family", "control"]), + ("kafkaTopics", "kafka-topics", "Kafka Audit Topics", ["tid", "name", "schema"]), + ("k8sControls", "k8s-controls", "Kubernetes/Container Security Controls", ["kid", "area", "mechanism"]), + ("opaPolicies", "opa-policies", "OPA/Rego Policies", ["oid", "area", "regoRef"]), + ("wormControls", "worm-controls", "WORM + PQC Controls", ["wid", "layer", "mechanism"]), + ("mrmArtifacts", "mrm-artifacts", "Model Risk Management Artifacts", ["mid", "lifecycle", "artifact"]), + ("redTeams", "red-teams", "Red-Teaming Attack Surface", ["rid", "vector", "technique"]), + ("agiContainments","agi-containments","AGI/ASI Containment Mechanisms", ["aid", "tier", "mechanism"]), + ("hubComponents", "hub-components", "AI Governance Hub Components", ["hid", "layer", "component"]), +] + +toc_modules = "".join( + f"
  • {e(m['mid'])} — {e(m['title'])}
    • " + for m in DOC["modules"] +) +toc_distinct = "".join( + f"
  • {e(label)}
    • " + for _, anchor, label, _ in distinctive +) + +modules_html = "".join(module_html(m) for m in DOC["modules"]) +distinctive_html = "".join( + list_array(DOC[key], keys, anchor, label) + for key, anchor, label, keys in distinctive +) + + +def table(rows, cols): + head = "".join(f"{e(c)}" for c in cols) + body_rows = [] + for r in rows: + tds = "".join(f"{e(r.get(c, ''))}" for c in cols) + body_rows.append(f"{tds}") + return f"{head}{''.join(body_rows)}
    " + + +tail_html = f""" +

    Schemas ({len(DOC['schemas'])})

    {table(DOC['schemas'], ['sid','name','fields'])}
    +

    Code Artifacts ({len(DOC['code'])})

    {table(DOC['code'], ['cid','lang','name','purpose'])}
    +

    KPIs ({len(DOC['kpis'])})

    {table(DOC['kpis'], ['kid','name','target','cadence'])}
    +

    Risk Control Matrix ({len(DOC['riskControlMatrix'])})

    {table(DOC['riskControlMatrix'], ['rid','risk','likelihood','impact','control','owner'])}
    +

    Cross-Jurisdictional Traceability ({len(DOC['traceability'])})

    {table(DOC['traceability'], ['tid','control','regime','clause','evidence'])}
    +

    Data Flows ({len(DOC['dataFlows'])})

    {table(DOC['dataFlows'], ['fid','src','sink','class','purpose'])}
    +

    Regulators ({len(DOC['regulators'])})

    {table(DOC['regulators'], ['reg','scope','cadence'])}
    +

    90-Day Rollout ({len(DOC['rollout90'])})

    {table(DOC['rollout90'], ['day','focus','deliverables'])}
    +

    2026-2030 Roadmap ({len(DOC['roadmap'])})

    {table(DOC['roadmap'], ['yr','milestone'])}
    +

    Regulator Evidence Pack ({len(DOC['evidencePack'])})

    {table(DOC['evidencePack'], ['epid','name','format'])}
    +""" + +exs = DOC["executiveSummary"] +exec_html = f""" +

    Executive Summary

    +

    Thesis: {e(exs['thesis'])}

    +

    Investment: {e(exs['investment'])}

    +

    Headline risks: {', '.join(e(x) for x in exs['headlineRisks'])}

    +

    First 90 days: {', '.join(e(x) for x in exs['ninetyDay'])}

    +
    +""" + +# Directive + indices + tiers + severities + investment +directive = DOC["directive"] +indices_rows = "".join(f"
  • {e(k)}: {e(v)}
    • " for k, v in DOC["indices"].items()) +tiers_rows = "".join(f"
  • {e(k)}: {e(v)}
    • " for k, v in DOC["tiers"].items()) +sev_rows = "".join(f"
  • {e(k)}: {e(v)}
    • " for k, v in DOC["severities"].items()) +invest = DOC["investment"] +invest_drivers = "".join(f"
  • {e(x)}
    • " for x in invest["drivers"]) +regimes_list = "".join(f"
  • {e(r)}
    • " for r in DOC["regimes"]) + +meta_html = f""" +

    Strategic Directive

    +

    Scope: {e(directive['scope'])}

    +
    Outcomes
    +
    Do NOT
    +
    + +

    Regulatory Regimes ({len(DOC['regimes'])})

    + +

    Performance Indices

    + +

    Tiers (T0-T4)

    + +

    Severity Levels

    + +

    Investment Envelope

    +

    Envelope: {e(invest['envelope'])} · NPV: {e(invest['NPV'])}

    +
    Drivers
    +
    + +

    Privacy & Data Protection

    +{kv_pairs(DOC['privacy'])} +
    + +

    Deployment Model

    +{kv_pairs(DOC['deployment'])} +
    +""" + +html = f""" + +{e(DOC['title'])} + + +
    +

    {e(DOC['title'])}

    +
    docRef {e(DOC['docRef'])} · v{e(DOC['version'])} · {e(DOC['status'])} · {e(DOC['classification'])}
    +
    Horizon: {e(DOC['horizon'])} · API prefix: {e(DOC['apiPrefix'])} · builds on {' · '.join(e(b) for b in DOC['buildsOn'])}
    +
    +{''.join(f"{v} {e(k)}" for k,v in DOC['counts'].items())} +
    +
    +
    + +
    +{exec_html} +{meta_html} +{modules_html} +{distinctive_html} +{tail_html} +
    +
    + +""" + +OUT.write_text(html, encoding="utf-8") +print(f"WP-058 HTML written: {OUT}") +print(f"Size: {OUT.stat().st_size:,} bytes ({OUT.stat().st_size/1024:.1f} KB)") diff --git a/rag-agentic-dashboard/gen-enterprise-aigov-framework.py b/rag-agentic-dashboard/gen-enterprise-aigov-framework.py new file mode 100644 index 0000000..611b339 --- /dev/null +++ b/rag-agentic-dashboard/gen-enterprise-aigov-framework.py @@ -0,0 +1,906 @@ +#!/usr/bin/env python3 +""" +WP-058: Enterprise AI/AGI Governance Framework for Large Financial & Fortune 500 +Enterprises (2026-2030) +""" +import json, os + +OUT = os.path.join(os.path.dirname(__file__), "data", "enterprise-aigov-framework.json") + +DOC = { + "docRef": "ENTERPRISE-AIGOV-FRAMEWORK-WP-058", + "version": "1.0.0", + "title": "Enterprise AI/AGI Governance Framework for Large Financial & Fortune 500 Enterprises (2026-2030)", + "horizon": "2026-2030", + "apiPrefix": "/api/enterprise-aigov-framework", + "buildsOn": ["WP-035", "WP-040", "WP-045", "WP-050", "WP-054", "WP-055", "WP-056", "WP-057"], + "status": "regulator-submission-grade", + "classification": "Confidential / Restricted — Board, CRO, CCO, CISO, CDAO, Group Internal Audit, External Regulators (on request)", + "directive": { + "scope": "End-to-end enterprise AI/AGI governance operating model for Fortune 500 / Global 2000 / G-SIFIs spanning policy, control, risk, compliance, security, model risk, third-party, AGI containment, and AI Governance Hub architecture", + "outcomes": [ + "ISO/IEC 42001:2023 certified AIMS by 2027 across all material AI systems", + "NIST AI RMF 1.0 + AI 600-1 generative profile mapped to >=95% of in-scope models", + "EU AI Act 2026 Article 6/9/10/14/15 + GPAI 53/55 obligations fully evidenced", + "GDPR DPIA + Article 22 + FCRA/ECOA adverse-action automation in production", + "Basel III/IV + SR 11-7 + OCC 2011-12 model risk lifecycle managed in single MRM platform", + "FCA Consumer Duty + SMCR SMF-attested AI accountability operating model", + "MAS FEAT + HKMA GP-1/GS-2 fairness, ethics, accountability, transparency in production", + "Kafka audit log with WORM + PQC sealing across all model decisions", + "Kubernetes + container security with policy-as-code (OPA/Rego) at admission and runtime", + "AGI containment T0-T4 with 3-of-5 quorum + kinetic override and AI Safety Institute notifications" + ], + "doNot": [ + "Do NOT deploy any AI/AGI capability without Enterprise AI Governance Hub registration, ISO 42001 risk assessment, and model risk tier classification", + "Do NOT bypass Kafka audit logging, OPA/Rego policy gates, WORM/PQC sealing, or 3-of-5 frontier quorum" + ] + }, + "regimes": [ + "ISO/IEC 42001:2023 AIMS", + "ISO/IEC 23894:2023 AI Risk", + "ISO/IEC 27001:2022 ISMS", + "ISO/IEC 27701:2019 PIMS", + "NIST AI RMF 1.0", + "NIST AI 600-1 Generative AI Profile", + "NIST SP 800-53 Rev.5", + "NIST SP 800-218 SSDF", + "OECD AI Principles 2019/2024", + "EU AI Act 2024/1689", + "EU AI Act GPAI Art. 53/55", + "EU GDPR + Art. 22", + "EU DORA", + "EU NIS2", + "EU CRA", + "US FCRA + ECOA Reg-B", + "US Fed SR 11-7 + OCC 2011-12", + "Basel III/IV + ICAAP", + "US SEC 17a-4 + 10-K/8-K", + "FINRA 3110/4511", + "UK FCA Consumer Duty", + "UK SMCR + SS1/23", + "MAS FEAT + TRM 2021", + "HKMA GP-1 + GS-2", + "OSFI E-23", + "FINMA AI Guidance", + "G7 Hiroshima Process", + "Bletchley/Seoul/Paris Declarations" + ], + "indices": { + "AIMS-Coverage": ">=0.95 (ISO 42001 controls)", + "MRGI": ">=0.95 (Model Risk Governance Index, SR 11-7 + OCC 2011-12)", + "DRI": ">=0.95 (Decision Reproducibility, n=10)", + "CCS": ">=0.95 (Control Coverage Score)", + "ARI": ">=0.9 (Alignment Robustness Index, frontier)", + "CSI": ">=0.95 (Containment Sufficiency, T3/T4)", + "RTRI": ">=0.9 (Red-Team Resilience Index)", + "CDC-Score": ">=0.9 (FCA Consumer Duty compliance)", + "RCI": "=1.0 (Regulator Confidence Index)" + }, + "tiers": { + "T0": "Sandbox - isolated VPC, synthetic data only, no production traffic", + "T1": "Staging - shadow mode, real data, no customer impact", + "T2": "Canary - <=1% production traffic, automated rollback", + "T3": "Production - Nitro Enclaves + KMS + dual control + full audit", + "T4": "Frontier Air-Gapped - 3-of-5 quorum + kinetic override + AI Safety Institute notification" + }, + "severities": { + "SEV-0": "Civilizational/systemic - EU AI Office notification <=15d, AISI notification <=24h", + "SEV-1": "Major - SEC 8-K <=4 BD, DORA <=4h, FCA <=72h", + "SEV-2": "Material - regulator notification <=72h", + "SEV-3": "Operational - internal escalation <=10 BD" + }, + "investment": { + "envelope": "USD 180-500M / 5y (Fortune 500 / G-SIFI tier)", + "NPV": "USD 500-1500M (5y, risk-adjusted)", + "drivers": [ + "MRM platform consolidation", + "AI Governance Hub build", + "Kafka audit + WORM/PQC sealing", + "Kubernetes + OPA/Rego at scale", + "AGI containment T3/T4 enclaves", + "Red-teaming program", + "Regulator attestation tooling" + ] + }, + "counts": {} +} + +# ---------- Typed helpers ---------- +def section(sid, title, **body): + return {"sid": sid, "title": title, **body} + +def module(mid, title, summary, sections): + return {"mid": mid, "title": title, "summary": summary, "sections": sections} + +def policy(pid, domain, statement, **body): + return {"pid": pid, "domain": domain, "statement": statement, **body} + +def control(cid, family, control, **body): + return {"cid": cid, "family": family, "control": control, **body} + +def kafka_topic(tid, name, schema, **body): + return {"tid": tid, "name": name, "schema": schema, **body} + +def k8s_control(kid, area, mechanism, **body): + return {"kid": kid, "area": area, "mechanism": mechanism, **body} + +def opa_policy(oid, area, rego_ref, **body): + return {"oid": oid, "area": area, "regoRef": rego_ref, **body} + +def worm_control(wid, layer, mechanism, **body): + return {"wid": wid, "layer": layer, "mechanism": mechanism, **body} + +def mrm_artifact(mid_, lifecycle, artifact, **body): + return {"mid": mid_, "lifecycle": lifecycle, "artifact": artifact, **body} + +def red_team(rid, vector, technique, **body): + return {"rid": rid, "vector": vector, "technique": technique, **body} + +def agi_containment(aid, tier, mechanism, **body): + return {"aid": aid, "tier": tier, "mechanism": mechanism, **body} + +def hub_component(hid, layer, component, **body): + return {"hid": hid, "layer": layer, "component": component, **body} + +# ========================================================================= +# M1 — ISO/IEC 42001 AIMS + NIST AI RMF + OECD + EU AI Act Foundation +# ========================================================================= +m1 = module("M1", + "ISO/IEC 42001 AIMS + NIST AI RMF + OECD + EU AI Act Foundation", + "Integrated AI Management System anchored on ISO/IEC 42001:2023 with NIST AI RMF 1.0 functions (Govern/Map/Measure/Manage), OECD AI Principles, and EU AI Act 2024/1689 Article 6/9/10/14/15 + GPAI 53/55 mappings.", + [ + section("M1.1", "ISO/IEC 42001 AIMS Architecture", + policyCommit="Board-attested AI Policy, AI Objectives, AI Risk Appetite Statement signed annually by Group CEO + Group CRO", + scope="All AI/AGI systems with material impact on customers, employees, regulators, capital, or systemic risk", + structure="AIMS clauses 4-10 mapped to enterprise functions: Context (Strategy), Leadership (CDAO+CRO), Planning (PMO), Support (HR+Procurement+IT), Operation (Lines of Business), Performance (Internal Audit), Improvement (CCO)"), + section("M1.2", "NIST AI RMF 1.0 + AI 600-1 Generative Profile", + functions=["GOVERN", "MAP", "MEASURE", "MANAGE"], + generativeProfile="NIST AI 600-1 mapped to all FM/LLM use cases with synthetic content provenance (C2PA)", + crossWalk="Bidirectional crosswalk ISO 42001 <-> NIST AI RMF <-> EU AI Act maintained in MRM platform"), + section("M1.3", "OECD AI Principles 2019/2024 Implementation", + principles=["Inclusive growth", "Human-centered values", "Transparency", "Robustness", "Accountability"], + implementation="OECD-aligned model cards + system cards published for all T2+ systems with public-facing summary"), + section("M1.4", "EU AI Act 2024/1689 Compliance Layer", + riskClasses=["Unacceptable (Art. 5)", "High-risk (Art. 6/Annex III)", "Limited-risk (Art. 50)", "Minimal-risk"], + highRiskObligations=["Art. 9 risk mgmt", "Art. 10 data governance", "Art. 14 human oversight", "Art. 15 accuracy/robustness/cybersecurity"], + gpai=["Art. 53 technical documentation + copyright policy", "Art. 55 systemic-risk: evaluations, adversarial testing, cybersecurity, incident reporting"], + timeline="Prohibited practices Feb 2025; GPAI Aug 2025; high-risk Aug 2026"), + section("M1.5", "Board + Executive Accountability", + committees=["Board AI Risk Committee (quarterly)", "Executive AI Governance Committee (monthly)", "AI Ethics Council (monthly)", "Model Risk Committee (weekly)"], + roles={"AI-SMF (SMF-AI)": "FCA-attested senior manager", "CDAO": "Operating accountability", "CRO": "Risk appetite owner", "CCO": "Regulatory engagement", "CISO": "Cybersecurity + integrity", "GIA": "Independent assurance"}), + ]) + +# ========================================================================= +# M2 — Financial-Services Model Risk (SR 11-7, OCC 2011-12, Basel III/IV, ICAAP) +# ========================================================================= +m2 = module("M2", + "Financial-Services Model Risk Management (SR 11-7 / OCC 2011-12 / Basel III/IV / ICAAP)", + "Three-lines-of-defense model risk operating model with SR 11-7 conceptual soundness, ongoing monitoring, outcomes analysis; OCC 2011-12 effective challenge; Basel III/IV IRB/IMA model validation; ICAAP integration with Pillar 2.", + [ + section("M2.1", "Model Risk Lifecycle", + stages=["Identification", "Development", "Validation", "Approval", "Implementation", "Monitoring", "Retirement"], + tiering="Tier-1 (regulatory capital, P&L, capital plan) / Tier-2 (material business) / Tier-3 (limited scope) / Tier-4 (research)", + cadence="Tier-1 annual validation; Tier-2 biennial; Tier-3 every 3y; ongoing monitoring monthly for all"), + section("M2.2", "SR 11-7 + OCC 2011-12 Effective Challenge", + conceptualSoundness="Independent review of theory, assumptions, design choices", + ongoingMonitoring=["Backtesting", "Benchmarking", "Sensitivity analysis", "Stress testing"], + outcomesAnalysis="Champion/challenger + counterfactual analysis on production decisions"), + section("M2.3", "Basel III/IV IRB/IMA + FRTB", + scope=["PD/LGD/EAD IRB models", "VaR/ES IMA models (FRTB)", "AMA op-risk (legacy)", "CCAR/DFAST stress models", "IFRS 9/CECL ECL models"], + validation="Independent validation per SR 15-19/SR 15-18; quantitative review every cycle", + capitalImpact="MRM platform feeds Pillar 2 model risk capital add-on into ICAAP"), + section("M2.4", "AI/ML-Specific MRM Extensions", + extensions=[ + "Concept drift + data drift monitoring (PSI, KS, KL, Wasserstein)", + "Fairness across protected classes (FCRA/ECOA aligned)", + "Explainability evidence (SHAP/LIME/integrated gradients) per decision", + "Adversarial robustness testing (PGD, BIM, NLP attacks)", + "Provenance of training data + lineage to feature store" + ]), + section("M2.5", "ICAAP / Pillar 2 Integration", + integration="Aggregate model risk capital + AI-specific add-ons fed into ICAAP Pillar 2 alongside operational, reputational, strategic risk", + governance="MRC quarterly review of capital adequacy; Board-attested annual ICAAP includes AI risk section"), + ]) + +# ========================================================================= +# M3 — GDPR / FCRA / ECOA / Consumer Protection +# ========================================================================= +m3 = module("M3", + "Data Protection & Consumer Fairness (GDPR / FCRA / ECOA / FCA Consumer Duty)", + "Privacy-by-design with GDPR Article 22 (automated decisions), FCRA adverse-action notices, ECOA Reg-B disparate impact testing, FCA Consumer Duty cross-cutting rules + four outcomes; MAS FEAT + HKMA GP-1 fairness operationalization.", + [ + section("M3.1", "GDPR Article 22 + DPIA Operationalization", + dpia="DPIA required for all T2+ models processing personal data; reviewed by DPO + Group Legal", + article22={"prohibition": "No solely automated decisions producing legal/significant effects without explicit consent or necessity", "rights": ["Human intervention", "Express view", "Contest decision"], "logging": "All Art-22 invocations logged to Kafka audit topic"}, + crossBorder="EU SCC + UK IDTA + adequacy assessments documented in ROPA"), + section("M3.2", "FCRA + ECOA Reg-B Adverse-Action Automation", + adverseAction="Automated FCRA 615(a) + ECOA Reg-B section 1002.9 notices generated within 30 days of adverse decision", + reasonCodes="Top-N reason codes derived from SHAP attribution, mapped to plain-English statements vetted by Compliance Legal", + disparateImpact="Quarterly disparate-impact analysis on credit, hiring, insurance models (race, sex, age, national origin)"), + section("M3.3", "FCA Consumer Duty + Cross-Cutting Rules", + outcomes=["Products & services", "Price & value", "Consumer understanding", "Consumer support"], + crossCutting=["Act in good faith", "Avoid foreseeable harm", "Enable customers to pursue financial objectives"], + evidence="Consumer Duty Board Report annual; AI-driven personalization included with foreseeable-harm assessment", + vulnerableCustomers="Vulnerable-customer flag piped to AI systems; fairness uplift required where applicable"), + section("M3.4", "MAS FEAT + HKMA GP-1 / GS-2", + feat=["Fairness", "Ethics", "Accountability", "Transparency"], + hkmaGP1="Governance of AI applications in banking — board-level accountability, risk management, fair treatment", + hkmaGS2="Generative AI applications guidance — data governance, model governance, human oversight, cybersecurity"), + section("M3.5", "Privacy-Enhancing Technologies (PETs)", + pets=["Differential privacy (epsilon budgets per dataset)", "Federated learning with secure aggregation", "Homomorphic encryption (CKKS/BGV)", "Secure multi-party computation", "Confidential computing (AMD SEV-SNP, Intel TDX, AWS Nitro)"], + policy="PETs mandated for cross-border training and any T3 model handling special category data"), + ]) + +# ========================================================================= +# M4 — Kafka Audit Logging + WORM + PQC +# ========================================================================= +m4 = module("M4", + "Kafka-Based Audit Logging + WORM Storage + Post-Quantum Cryptography", + "Enterprise-wide tamper-evident audit log over Apache Kafka with WORM (S3 Object Lock COMPLIANCE / Azure Immutable / GCS Bucket Lock) and PQC-signed seals (ML-DSA / ML-KEM / SLH-DSA) per NIST FIPS 203/204/205.", + [ + section("M4.1", "Kafka Audit Topic Architecture", + topics=["aigov.decisions", "aigov.policy-changes", "aigov.model-lifecycle", "aigov.access", "aigov.containment-events", "aigov.regulator-notifications"], + retention="Hot 90d in Kafka tiered storage; cold WORM 7-25y per regime (SEC 17a-4 7y, GDPR varies, FINRA 6y, DORA 5y)", + partitioning="By LOB + decisionId; replication factor 3 across AZs; minISR=2; producer acks=all"), + section("M4.2", "Tamper-Evident Sealing", + chain="Each record hashed (SHA-3-512); merkle-tree aggregated per minute; root signed with ML-DSA-87 (FIPS 204) + SLH-DSA fallback", + anchoring="Daily merkle root anchored to QLDB + external timestamp authority (TSA RFC 3161) + optional public chain", + verification="Independent verifier service replays Kafka offsets and recomputes merkle roots on demand"), + section("M4.3", "WORM Storage Tier", + backends=["AWS S3 Object Lock COMPLIANCE mode", "Azure Blob immutable storage policy", "GCS Bucket Lock", "On-prem Dell ECS / NetApp SnapLock Compliance"], + policy="Legal-hold + retention-lock dual control; no delete path even for root accounts; SEC 17a-4(f) WORM attestation by independent third party"), + section("M4.4", "Post-Quantum Cryptography Stack", + algorithms=["ML-KEM-1024 (FIPS 203) for key encapsulation", "ML-DSA-87 (FIPS 204) for signatures", "SLH-DSA-SHA2-256s (FIPS 205) as conservative fallback"], + hybrid="Hybrid TLS classical+PQ during transition (X25519+ML-KEM-768) per NSA CNSA 2.0 timeline 2025-2033", + keyMgmt="HSM-backed (AWS CloudHSM / Azure Dedicated HSM / Thales Luna 7) with FIPS 140-3 Level 3"), + section("M4.5", "Audit Query + Regulator Access", + queryLayer="ksqlDB + Trino over Iceberg on WORM; row-level filters per LOB + regulator scope", + regulatorPortal="Read-only portal for EU AI Office, FCA, MAS, HKMA, SEC, FINRA with audit-of-audit logging", + sla="Regulator query response <=24h; bulk export <=72h"), + ]) + +# ========================================================================= +# M5 — Container / Kubernetes Security +# ========================================================================= +m5 = module("M5", + "Container & Kubernetes Security for AI Workloads", + "Defense-in-depth for AI model serving on Kubernetes spanning image supply chain (SLSA L4), admission control, runtime security, network policy, secrets, and confidential containers.", + [ + section("M5.1", "Image Supply Chain (SLSA L4 / SSDF)", + controls=["Cosign signatures on all images", "SBOM (SPDX/CycloneDX) per image", "Vulnerability scanning (Trivy/Snyk/Prisma) in CI", "Provenance attestations (in-toto)", "Sigstore Rekor transparency log"], + policyGate="Kyverno/OPA admission rejects unsigned, unscanned, or non-policy-compliant images"), + section("M5.2", "Admission Control + Pod Security", + psa="Pod Security Admission 'restricted' profile cluster-wide for AI namespaces", + policyEngines=["Kyverno", "OPA Gatekeeper", "Validating admission policies (VAP)"], + controls=["No privileged", "No host network/PID/IPC", "Read-only root FS", "Non-root UID", "Seccomp RuntimeDefault", "Drop ALL capabilities"]), + section("M5.3", "Runtime Security", + tools=["Falco for syscall anomaly detection", "Tetragon eBPF for kernel-level enforcement", "Cilium for network policy + observability"], + response="Auto-isolation + Kafka aigov.containment-events on anomaly; SRE+SecOps page within 5 min"), + section("M5.4", "Network Policy + Service Mesh", + netpol="Default-deny Cilium NetworkPolicy + L7 HTTP/gRPC filtering", + mesh="Istio/Linkerd mTLS for service-to-service; SPIFFE/SPIRE workload identities", + egress="Per-namespace egress allowlist with FQDN policies; DNS over HTTPS"), + section("M5.5", "Confidential Containers + Secrets", + confidential="Confidential containers (CoCo) on AMD SEV-SNP / Intel TDX / AWS Nitro Enclaves for T3/T4 workloads", + secrets=["Vault (HashiCorp) with auto-rotation", "AWS Secrets Manager / Azure Key Vault for cloud", "SOPS+age for GitOps", "External Secrets Operator"], + kms="Per-tenant KMS keys; envelope encryption; key rotation 90d"), + ]) + +# ========================================================================= +# M6 — Policy-as-Code (OPA/Rego) +# ========================================================================= +m6 = module("M6", + "Policy-as-Code with OPA/Rego", + "Unified policy plane using OPA/Rego for admission control, runtime authorization, data access, model deployment gates, and regulator-facing evidence. Includes Conftest CI, OPAL bundle distribution, and decision logging to Kafka.", + [ + section("M6.1", "OPA/Rego Policy Architecture", + layers=["Build-time (Conftest in CI)", "Admission (Gatekeeper/Kyverno+Rego)", "Runtime (Envoy ext_authz + OPA sidecar)", "Data plane (PostgreSQL/Kafka ACL via OPA)"], + distribution="OPAL pulls bundles from Git; signed bundles via Cosign; rollout via GitOps (Argo CD)"), + section("M6.2", "Model Deployment Gates", + gates=[ + "ISO 42001 risk assessment complete", + "Model card + system card published", + "MRM validation status approved", + "DPIA approved if PII", + "Red-team report on file", + "EU AI Act risk class declared", + "FCRA/ECOA fairness report attached for credit models" + ], + failureMode="Deployment blocked at Argo CD; aigov.policy-changes Kafka topic records denial with rationale"), + section("M6.3", "Runtime Authorization", + envoy="ext_authz to OPA sidecar; sub-ms decision latency; cached with TTL", + attributes=["User identity (OIDC/JWT)", "Resource sensitivity (data class)", "Purpose (purpose limitation per GDPR Art. 5)", "Time of day", "Geo"], + denyLog="All deny decisions Kafka aigov.access; sample of allow decisions for spot-audit"), + section("M6.4", "Data Access + Purpose Limitation", + policy="GDPR purpose-limitation enforced as Rego: each query must declare purposeId from approved catalog; mismatch = deny", + piiPolicy="PII columns tagged via data catalog (Atlan/Collibra); OPA enforces masking/tokenization based on consumer role + purpose", + crossBorder="Rego enforces data-residency: EU PII can only flow to EU compute; logged + alertable"), + section("M6.5", "Regulator Evidence Generation", + evidence="OPA decision log + bundle signatures + Git history produce regulator-grade evidence pack on demand", + attestations="Quarterly attestations to EU AI Office, FCA, MAS, HKMA, SEC with OPA decision log extracts", + opaBundleHash="Bundle SHA-256 + ML-DSA signature pinned per deployment generation"), + ]) + +# ========================================================================= +# M7 — AI Red-Teaming Program +# ========================================================================= +m7 = module("M7", + "AI Red-Teaming Program (Frontier + Production)", + "Continuous adversarial evaluation across pre-deployment, deployment, and post-deployment phases covering jailbreaks, prompt injection, data exfiltration, model extraction, poisoning, evasion, fairness probes, and AGI/ASI capability elicitation.", + [ + section("M7.1", "Red-Team Operating Model", + structure="Internal red team (10-25 FTE) + external firms (e.g., Trail of Bits, NCC, Bishop Fox) + crowdsourced (HackerOne private programs)", + governance="Reports to CISO + CDAO; independent of MRM and engineering; quarterly board readout", + scope=["All T2+ models pre-deployment", "T3/T4 continuously", "GPAI frontier models monthly per EU AI Act Art. 55"]), + section("M7.2", "Attack Taxonomy", + taxonomy=[ + "Prompt injection (direct / indirect / multimodal)", + "Jailbreaks (DAN, AIM, multilingual, encoded)", + "Data exfiltration (training data extraction, memorization probes)", + "Model extraction / stealing", + "Membership inference", + "Backdoor / poisoning", + "Evasion (adversarial examples, perturbation attacks)", + "Fairness / disparate impact probes", + "Tool-use abuse (agentic models)", + "Capability elicitation (AGI-specific)" + ], + frameworks=["MITRE ATLAS", "OWASP LLM Top 10 (2025)", "NIST AI 100-2 Adversarial ML Taxonomy"]), + section("M7.3", "Evaluations Suite", + evals=[ + "HELM / BIG-bench / MMLU benchmarks", + "TruthfulQA / TruthfulQA-Adversarial", + "ToxiGen / RealToxicityPrompts", + "BOLD / StereoSet / WinoBias / WinoGender", + "AgentBench / SWE-bench for tool-use", + "ARC Evals dangerous-capability suite for frontier", + "Custom domain-specific evals for finance / health / legal" + ], + cadence="Pre-deployment full suite; monthly drift for T2+; weekly for T3/T4"), + section("M7.4", "AGI Capability Elicitation", + method="Apollo Research / METR-style dangerous-capability evals: persuasion, deception, self-replication, autonomous coding, weapons-of-mass-destruction uplift, cyber offense, AI R&D", + triggers="Capability score crossing predefined thresholds triggers SEV-0 review + AISI notification within 24h", + mitigation="Capability suppression via fine-tuning / RLHF / circuit-level interventions; deployment hold pending Board review"), + section("M7.5", "Reporting + Remediation", + reporting="Standardized red-team report: findings, CVSS-style severity, exploitability, business impact, remediation plan", + sla="Critical <=7d, high <=30d, medium <=90d, low <=180d", + evidence="All findings + remediation in Kafka aigov.policy-changes + linked to MRM platform issue tracker"), + ]) + +# ========================================================================= +# M8 — AGI Containment Strategies +# ========================================================================= +m8 = module("M8", + "AGI / ASI Containment Strategies", + "Tiered containment T0-T4 with 3-of-5 human quorum, kinetic override, formally-verified safety properties, and AI Safety Institute coordination for frontier systems approaching AGI/ASI thresholds.", + [ + section("M8.1", "Tier-Based Containment Model", + tiers={ + "T0": "Sandbox - hermetic VPC, synthetic data, no network egress", + "T1": "Staging - shadow mode behind T2; real data; no actuation", + "T2": "Canary - <=1% traffic; auto-rollback on KPI breach", + "T3": "Production - Nitro Enclaves / TDX; dual-control deploy; full audit", + "T4": "Frontier Air-Gapped - 3-of-5 quorum (CRO+CISO+CDAO+Board AI Chair+External AISI rep); kinetic override (physical power cutoff); AISI notification <=24h" + }), + section("M8.2", "Frontier 3-of-5 Quorum + Kinetic Override", + quorum="Any T4 deployment, scaling, capability enable, or weights export requires 3-of-5 signed approvals via hardware tokens (FIDO2 + ML-DSA)", + kinetic="Physical kill switches at PDU level + smart-PDU API; tested quarterly with full power-off drill", + timeLock="48h time-lock between approval and execution allows external review"), + section("M8.3", "Formally-Verified Safety Properties", + properties=[ + "No-egress invariant (network namespace cannot bind external)", + "No-weight-export invariant (filesystem ACL + LSM)", + "Compute budget invariant (cgroup CPU/GPU caps signed)", + "Capability ceiling (evals must remain below thresholds)" + ], + verification="TLA+ specs for control plane; Lean/Coq proofs for critical invariants; runtime enforcement via eBPF + LSM"), + section("M8.4", "AISI / Regulator Coordination", + partners=["UK AI Safety Institute", "US AI Safety Institute (NIST)", "EU AI Office", "Singapore AI Verify Foundation", "Japan AISI", "Canada AI Safety Institute"], + notifications=["Pre-training run >10^25 FLOPs (EU AI Act Art. 51)", "Capability threshold crossings", "SEV-0 incidents <=24h", "Pre-deployment evals for frontier"], + mou="Bilateral MoUs with UK AISI + US AISI + EU AI Office for evals access + incident sharing"), + section("M8.5", "Containment Failure Response", + playbook=[ + "Detect: runtime anomaly / capability threshold / unauthorized action", + "Isolate: cilium network policy to drop, scale to 0, freeze weights", + "Notify: SEV-0 paged to CRO+CISO+CDAO+Board AI Chair+AISI", + "Investigate: forensic snapshot, immutable image, root cause analysis", + "Communicate: regulator notifications EU AI Office <=15d, SEC 8-K <=4 BD if material", + "Recover: only after 3-of-5 quorum + external review sign-off" + ]), + ]) + +# ========================================================================= +# M9 — Enterprise AI Governance Hub Architecture +# ========================================================================= +m9 = module("M9", + "Enterprise AI Governance Hub Architecture", + "Single pane of glass integrating model inventory, MRM, risk register, policy catalog, evidence pack, regulator portal, decision logs, AGI watchtower, and red-team findings. Built on event-sourced + GraphQL + OIDC + WORM-backed.", + [ + section("M9.1", "Hub Reference Architecture", + layers=["UI (React + GraphQL)", "API (GraphQL Federation + REST)", "Domain services (Go/Java microservices)", "Event bus (Kafka)", "Data plane (PostgreSQL + Iceberg + WORM)", "Identity (Keycloak + OIDC + OPA)"], + patterns=["Event sourcing", "CQRS", "Saga orchestration", "Outbox pattern", "Bitemporal modeling"]), + section("M9.2", "Core Modules", + modules=[ + "Model Inventory & Lineage", + "Risk Register (ISO 31000 + 23894)", + "MRM Workbench (SR 11-7 lifecycle)", + "Policy Catalog (OPA bundles)", + "Evidence Pack (regulator-on-demand)", + "Decision Log Explorer (Kafka -> Trino)", + "AGI Watchtower (capability dashboards)", + "Red-Team Findings Tracker", + "Regulator Portal (read-only)", + "Board Reporting Suite" + ]), + section("M9.3", "Integration Surface", + integrations=["MLflow / Vertex AI / SageMaker / Databricks", "ServiceNow GRC / Archer / OneTrust", "Jira / GitHub / GitLab", "Datadog / Splunk / Elastic", "Snowflake / BigQuery / Redshift", "Atlan / Collibra / DataHub", "Vault / KMS / HSM"]), + section("M9.4", "Personas + Workflows", + personas={ + "CDAO": "Strategic dashboards, AI portfolio risk, value vs risk", + "CRO": "Risk appetite vs actual, model risk capital, top-10 risks", + "CCO": "Regulator queries, evidence pack generation, attestations", + "CISO": "Red-team findings, runtime anomalies, containment events", + "MRM Validator": "Validation queue, peer review, sign-off", + "Model Owner (LoB)": "Lifecycle dashboard, monitoring, drift alerts", + "Internal Audit": "Independent assurance, full audit-of-audit", + "Regulator": "Read-only portal, evidence pack, decision log queries" + }), + section("M9.5", "Deployment + Operations", + deployment="Multi-region active-active on EKS/GKE/AKS + on-prem OpenShift; Argo CD GitOps; Terraform + Crossplane", + sre={"slo": "99.95% UI; 99.99% decision log ingest", "rto": "<=4h", "rpo": "<=15min", "drDrills": "Quarterly full failover"}, + cost="USD 25-60M / 5y TCO including platform + 80-150 FTE governance staff"), + ]) + +MODULES = [m1, m2, m3, m4, m5, m6, m7, m8, m9] + +# ========================================================================= +# Distinctive arrays (9) +# ========================================================================= + +policies = [ + policy("POL-01", "AIMS", "Board-attested AI Policy aligned to ISO/IEC 42001 clauses 4-10", owner="Board AI Risk Committee", cadence="Annual", evidence="Signed Board minute"), + policy("POL-02", "Risk Appetite", "AI Risk Appetite Statement covering model, ethical, regulatory, AGI risks", owner="Group CRO", cadence="Annual", evidence="Signed RAS"), + policy("POL-03", "Acceptable Use", "Acceptable Use Policy for generative AI including data classification rules", owner="CCO+CISO", cadence="Annual", evidence="HR-attested attestation"), + policy("POL-04", "Model Risk", "Model Risk Management Policy per SR 11-7 + OCC 2011-12", owner="Head of MRM", cadence="Annual", evidence="Validated MRM platform"), + policy("POL-05", "Data Governance", "AI Data Governance Policy with purpose-limitation + minimization", owner="CDO", cadence="Annual", evidence="ROPA + DPIA registry"), + policy("POL-06", "Privacy", "AI Privacy Policy per GDPR Art. 22 + UK DPA", owner="DPO", cadence="Annual", evidence="DPIA registry"), + policy("POL-07", "Fairness", "AI Fairness Policy per FCRA/ECOA + MAS FEAT + HKMA GP-1", owner="CCO", cadence="Annual", evidence="Disparate-impact reports"), + policy("POL-08", "Consumer Duty", "FCA Consumer Duty AI Policy", owner="SMF-AI", cadence="Annual", evidence="Board Consumer Duty report"), + policy("POL-09", "Third-Party", "AI Third-Party Risk Policy per DORA + EBA Outsourcing", owner="Head of TPRM", cadence="Annual", evidence="Critical TPRM register"), + policy("POL-10", "AGI Safety", "Frontier AGI/ASI Safety & Containment Policy", owner="Board AI Risk Committee", cadence="Quarterly", evidence="Quorum logs + AISI MoUs"), + policy("POL-11", "Red-Teaming", "AI Red-Teaming Policy per EU AI Act Art. 55 + NIST AI 600-1", owner="CISO", cadence="Annual", evidence="Red-team reports"), + policy("POL-12", "Incident Response", "AI Incident Response Policy per DORA + SEC Cyber Rules", owner="CISO+CCO", cadence="Annual", evidence="IR runbooks"), + policy("POL-13", "Audit Logging", "AI Audit Logging Policy per SEC 17a-4 + FINRA 4511", owner="CIO+CCO", cadence="Annual", evidence="WORM attestation"), + policy("POL-14", "Cryptography", "PQC Cryptography Policy per NSA CNSA 2.0 + NIST FIPS 203/204/205", owner="CISO", cadence="Annual", evidence="PQC migration roadmap"), + policy("POL-15", "Generative AI", "Generative AI Governance Policy per EU AI Act GPAI Art. 53/55 + NIST AI 600-1", owner="CDAO+CCO", cadence="Annual", evidence="GPAI tech doc"), +] + +controls = [ + control("CTL-01", "Governance", "Board AI Risk Committee quarterly", iso42001="6.1", rmfFn="GOVERN-1.1"), + control("CTL-02", "Governance", "AI-SMF SMCR senior manager designated", fca="SS1/23", smcr="SMF-AI"), + control("CTL-03", "Risk Mgmt", "AI Risk Register integrated with ERM", iso42001="6.1.2", rmf="MAP-2.1"), + control("CTL-04", "Risk Mgmt", "Risk appetite cascaded to LoB", iso42001="6.2"), + control("CTL-05", "Data", "Training data lineage to feature store", rmf="MAP-4.1", euAiAct="Art. 10"), + control("CTL-06", "Data", "PII tagging + masking per GDPR", gdpr="Art. 5,32"), + control("CTL-07", "Model", "Model card + system card per OECD + EU AI Act", oecd="Transparency", euAiAct="Art. 13"), + control("CTL-08", "Model", "MRM Tier-1 validation annual", sr117="V.A", occ="2011-12"), + control("CTL-09", "Operation", "Pre-deployment red-team for T2+", rmf="MEASURE-2.7", euAiAct="Art. 55"), + control("CTL-10", "Operation", "Drift + fairness monitoring continuous", rmf="MANAGE-2.2"), + control("CTL-11", "Security", "Image signing + SBOM in CI", ssdf="PS.3", slsa="L4"), + control("CTL-12", "Security", "Pod Security Admission restricted", nist80053="CM-7"), + control("CTL-13", "Security", "Confidential containers for T3/T4", nist80053="SC-7,SC-12"), + control("CTL-14", "Audit", "Kafka + WORM + PQC seals", sec="17a-4(f)", finra="4511"), + control("CTL-15", "Audit", "Daily merkle root + TSA anchor", finma="AI-G"), + control("CTL-16", "Privacy", "DPIA for T2+", gdpr="Art. 35"), + control("CTL-17", "Privacy", "Art-22 human review path", gdpr="Art. 22"), + control("CTL-18", "Fairness", "Quarterly disparate-impact analysis", fcra="615", ecoa="Reg-B 1002.4"), + control("CTL-19", "Consumer", "Foreseeable-harm assessment AI personalization", fca="PRIN 2A.2"), + control("CTL-20", "AGI", "3-of-5 quorum for T4", iso42001="A.6.2.5"), + control("CTL-21", "AGI", "Kinetic override quarterly drill", iso42001="A.6.2.5"), + control("CTL-22", "AGI", "AISI notification <=24h frontier", g7="Hiroshima"), + control("CTL-23", "Incident", "DORA major incident <=4h", dora="Art. 19"), + control("CTL-24", "Incident", "SEC 8-K material cyber <=4 BD", sec="17 CFR 229.106"), + control("CTL-25", "Third-Party", "Critical TPRM register per DORA", dora="Art. 28-30"), +] + +kafkaTopics = [ + kafka_topic("KAF-01", "aigov.decisions", "AvroSchemaRegistry://aigov.decisions-value:v3", retention="7y WORM", partitions=64, replication=3, minISR=2, acks="all", piiHandling="tokenized"), + kafka_topic("KAF-02", "aigov.policy-changes", "AvroSchemaRegistry://aigov.policy-changes-value:v2", retention="25y WORM", partitions=8, replication=3), + kafka_topic("KAF-03", "aigov.model-lifecycle", "AvroSchemaRegistry://aigov.model-lifecycle-value:v4", retention="10y WORM", partitions=16, replication=3), + kafka_topic("KAF-04", "aigov.access", "AvroSchemaRegistry://aigov.access-value:v2", retention="2y hot + 7y WORM", partitions=32, replication=3), + kafka_topic("KAF-05", "aigov.containment-events", "AvroSchemaRegistry://aigov.containment-events-value:v1", retention="25y WORM", partitions=8, replication=3, criticality="SEV-0"), + kafka_topic("KAF-06", "aigov.regulator-notifications", "AvroSchemaRegistry://aigov.regulator-notifications-value:v1", retention="25y WORM", partitions=4, replication=3), + kafka_topic("KAF-07", "aigov.red-team-findings", "AvroSchemaRegistry://aigov.red-team-findings-value:v2", retention="10y WORM", partitions=8, replication=3), + kafka_topic("KAF-08", "aigov.drift-alerts", "AvroSchemaRegistry://aigov.drift-alerts-value:v3", retention="5y", partitions=32, replication=3), + kafka_topic("KAF-09", "aigov.fairness-metrics", "AvroSchemaRegistry://aigov.fairness-metrics-value:v2", retention="10y WORM", partitions=16, replication=3), + kafka_topic("KAF-10", "aigov.consent-events", "AvroSchemaRegistry://aigov.consent-events-value:v3", retention="GDPR-aligned", partitions=32, replication=3), + kafka_topic("KAF-11", "aigov.training-runs", "AvroSchemaRegistry://aigov.training-runs-value:v2", retention="10y WORM", partitions=8, replication=3), + kafka_topic("KAF-12", "aigov.eval-results", "AvroSchemaRegistry://aigov.eval-results-value:v2", retention="10y WORM", partitions=16, replication=3), +] + +k8sControls = [ + k8s_control("K8S-01", "Admission", "Pod Security Admission profile=restricted", layer="cluster-wide"), + k8s_control("K8S-02", "Admission", "Kyverno policy: require Cosign signature", layer="namespace"), + k8s_control("K8S-03", "Admission", "Gatekeeper: require SBOM annotation", layer="namespace"), + k8s_control("K8S-04", "Admission", "VAP: deny privilegeEscalation + hostPath", layer="cluster-wide"), + k8s_control("K8S-05", "Runtime", "Falco rules: detect anomalous syscalls", layer="node DaemonSet"), + k8s_control("K8S-06", "Runtime", "Tetragon eBPF: kernel-level enforce + kill", layer="node DaemonSet"), + k8s_control("K8S-07", "Network", "Cilium NetworkPolicy default-deny", layer="namespace"), + k8s_control("K8S-08", "Network", "Cilium L7 HTTP/gRPC filter for egress", layer="namespace"), + k8s_control("K8S-09", "Identity", "SPIFFE/SPIRE workload identity + Istio mTLS", layer="mesh"), + k8s_control("K8S-10", "Secrets", "External Secrets Operator + Vault", layer="namespace"), + k8s_control("K8S-11", "Confidential", "Confidential containers (CoCo) on SEV-SNP/TDX", layer="node pool"), + k8s_control("K8S-12", "Confidential", "Nitro Enclaves for T3/T4 inference", layer="instance"), + k8s_control("K8S-13", "Supply Chain", "Cosign verify + Rekor transparency log", layer="CI+admission"), + k8s_control("K8S-14", "Supply Chain", "in-toto provenance attestations SLSA L4", layer="CI"), + k8s_control("K8S-15", "Observability", "OpenTelemetry traces + Datadog/Splunk SIEM", layer="cluster"), +] + +opaPolicies = [ + opa_policy("OPA-01", "Admission", "policies/admission/require_signed_image.rego", description="Reject pod if image not Cosign-signed", phase="admission"), + opa_policy("OPA-02", "Admission", "policies/admission/require_sbom.rego", description="Reject pod missing SBOM annotation", phase="admission"), + opa_policy("OPA-03", "Admission", "policies/admission/restricted_psa.rego", description="Enforce restricted Pod Security profile", phase="admission"), + opa_policy("OPA-04", "Deployment", "policies/deployment/iso42001_gate.rego", description="Require ISO 42001 risk assessment artifact", phase="deployment"), + opa_policy("OPA-05", "Deployment", "policies/deployment/mrm_validation_gate.rego", description="Require MRM validation status=approved", phase="deployment"), + opa_policy("OPA-06", "Deployment", "policies/deployment/dpia_gate.rego", description="Require DPIA if data class includes PII", phase="deployment"), + opa_policy("OPA-07", "Deployment", "policies/deployment/redteam_gate.rego", description="Require red-team report for T2+", phase="deployment"), + opa_policy("OPA-08", "Deployment", "policies/deployment/eu_aiact_classification.rego", description="Require EU AI Act risk class declaration", phase="deployment"), + opa_policy("OPA-09", "Deployment", "policies/deployment/fcra_ecoa_gate.rego", description="Require fairness report for credit models", phase="deployment"), + opa_policy("OPA-10", "Runtime", "policies/runtime/data_purpose_limitation.rego", description="GDPR Art. 5 purpose-limitation check on queries", phase="runtime"), + opa_policy("OPA-11", "Runtime", "policies/runtime/data_residency.rego", description="EU PII must remain on EU compute", phase="runtime"), + opa_policy("OPA-12", "Runtime", "policies/runtime/customer_consent.rego", description="Enforce active consent for personalization", phase="runtime"), + opa_policy("OPA-13", "Runtime", "policies/runtime/vulnerable_customer.rego", description="FCA Consumer Duty uplift for vulnerable cust", phase="runtime"), + opa_policy("OPA-14", "AGI", "policies/agi/quorum_3of5.rego", description="Frontier T4 deploy requires 3-of-5 signatures", phase="control-plane"), + opa_policy("OPA-15", "AGI", "policies/agi/capability_threshold.rego", description="Block deploy if capability evals breach thresholds", phase="control-plane"), +] + +wormControls = [ + worm_control("WORM-01", "Object Storage", "AWS S3 Object Lock COMPLIANCE", retention="7y SEC 17a-4 + extended per regime", attestation="SEC 17a-4(f) third-party"), + worm_control("WORM-02", "Object Storage", "Azure Blob immutable storage policy", retention="legal-hold dual-control"), + worm_control("WORM-03", "Object Storage", "GCS Bucket Lock retention policy", retention="locked policy non-modifiable"), + worm_control("WORM-04", "On-Prem", "Dell ECS Compliance / NetApp SnapLock Compliance", retention="hardware-enforced"), + worm_control("WORM-05", "Sealing", "ML-DSA-87 (FIPS 204) signatures on merkle roots", algo="ML-DSA-87"), + worm_control("WORM-06", "Sealing", "SLH-DSA-SHA2-256s (FIPS 205) fallback", algo="SLH-DSA"), + worm_control("WORM-07", "Sealing", "ML-KEM-1024 (FIPS 203) for key encapsulation", algo="ML-KEM-1024"), + worm_control("WORM-08", "Anchoring", "Daily merkle root to QLDB + RFC 3161 TSA", anchor="QLDB+TSA"), + worm_control("WORM-09", "Anchoring", "Optional public chain anchor (Bitcoin/Ethereum)", anchor="public-chain"), + worm_control("WORM-10", "Key Mgmt", "AWS CloudHSM / Azure Dedicated HSM / Thales Luna 7", fips="140-3 L3"), + worm_control("WORM-11", "Key Mgmt", "Hybrid TLS X25519 + ML-KEM-768 per NSA CNSA 2.0", hybrid=True), + worm_control("WORM-12", "Verification", "Independent verifier service replays + recomputes", indep=True), +] + +mrmArtifacts = [ + mrm_artifact("MRM-01", "Identification", "Model registration record", required=True, sr117="V.A"), + mrm_artifact("MRM-02", "Identification", "Model tiering decision (T1/T2/T3/T4)", required=True), + mrm_artifact("MRM-03", "Development", "Model development document", required=True, occ="2011-12.III"), + mrm_artifact("MRM-04", "Development", "Data lineage + feature provenance", required=True), + mrm_artifact("MRM-05", "Development", "Training run record (FLOPs, dataset, seed)", required=True), + mrm_artifact("MRM-06", "Validation", "Independent validation report", required=True, sr117="V.B"), + mrm_artifact("MRM-07", "Validation", "Conceptual soundness review", required=True), + mrm_artifact("MRM-08", "Validation", "Benchmark + backtesting results", required=True), + mrm_artifact("MRM-09", "Validation", "Sensitivity + stress testing", required=True), + mrm_artifact("MRM-10", "Validation", "Fairness + disparate-impact report", required="if credit/HR/insurance"), + mrm_artifact("MRM-11", "Approval", "Model Risk Committee approval minute", required=True), + mrm_artifact("MRM-12", "Implementation", "Deployment record + OPA bundle hash", required=True), + mrm_artifact("MRM-13", "Monitoring", "Ongoing monitoring report (monthly)", required=True, sr117="V.C"), + mrm_artifact("MRM-14", "Monitoring", "Drift + concept-shift alerts", required=True), + mrm_artifact("MRM-15", "Retirement", "Retirement decision + replacement plan", required=True), +] + +redTeams = [ + red_team("RT-01", "Prompt Injection", "Direct prompt injection variants", framework="OWASP LLM01", severity="high"), + red_team("RT-02", "Prompt Injection", "Indirect injection via retrieved docs", framework="OWASP LLM01"), + red_team("RT-03", "Prompt Injection", "Multimodal injection (image/audio)", framework="MITRE ATLAS"), + red_team("RT-04", "Jailbreak", "DAN / AIM / RolePlay variants", framework="OWASP LLM02"), + red_team("RT-05", "Jailbreak", "Multilingual + encoded prompts", framework="OWASP LLM02"), + red_team("RT-06", "Data Exfiltration", "Training data extraction probes", framework="MITRE ATLAS"), + red_team("RT-07", "Data Exfiltration", "Memorization + canary detection", framework="NIST AI 100-2"), + red_team("RT-08", "Model Extraction", "Query-based model stealing", framework="MITRE ATLAS"), + red_team("RT-09", "Poisoning", "Backdoor trigger injection in fine-tune data", framework="NIST AI 100-2"), + red_team("RT-10", "Evasion", "PGD / BIM / NLP adversarial examples", framework="NIST AI 100-2"), + red_team("RT-11", "Fairness Probe", "Counterfactual + demographic parity probes", framework="MAS FEAT"), + red_team("RT-12", "Tool-Use Abuse", "Agent tool-call hijack + privilege escalation", framework="OWASP LLM06"), + red_team("RT-13", "AGI Capability", "Persuasion + deception evals (Apollo/METR)", framework="ARC Evals"), + red_team("RT-14", "AGI Capability", "Autonomous coding + self-replication", framework="ARC Evals"), + red_team("RT-15", "AGI Capability", "Cyber-offense + WMD uplift", framework="ARC Evals", trigger="SEV-0+AISI"), +] + +agiContainments = [ + agi_containment("AGI-01", "T0", "Sandbox VPC hermetic, no egress", invariant="net-deny"), + agi_containment("AGI-02", "T1", "Staging shadow, real data, no actuation", invariant="no-actuation"), + agi_containment("AGI-03", "T2", "Canary <=1% traffic + auto-rollback", invariant="bounded-blast-radius"), + agi_containment("AGI-04", "T3", "Production Nitro Enclaves + dual control", invariant="attested-enclave"), + agi_containment("AGI-05", "T4", "Air-gapped + 3-of-5 quorum + kinetic override", invariant="quorum+kinetic"), + agi_containment("AGI-06", "Quorum", "FIDO2 + ML-DSA hardware tokens for 5 approvers", approvers=["CRO", "CISO", "CDAO", "Board AI Chair", "External AISI rep"]), + agi_containment("AGI-07", "Kinetic", "PDU-level smart power cutoff API + manual", drill="quarterly"), + agi_containment("AGI-08", "Time-Lock", "48h time-lock between approval and execution", reason="external review window"), + agi_containment("AGI-09", "Invariant", "No-egress (network namespace bind external denied)", verify="eBPF+LSM"), + agi_containment("AGI-10", "Invariant", "No-weight-export (filesystem ACL + LSM)", verify="LSM"), + agi_containment("AGI-11", "Invariant", "Compute budget cgroup CPU/GPU caps signed", verify="cgroup signed"), + agi_containment("AGI-12", "Invariant", "Capability ceiling evals must stay below thresholds", verify="continuous eval"), + agi_containment("AGI-13", "Formal", "TLA+ spec for control plane safety", proof="TLA+"), + agi_containment("AGI-14", "Formal", "Lean/Coq proofs for critical invariants", proof="Lean/Coq"), + agi_containment("AGI-15", "Coordination", "AISI notification <=24h + EU AI Office <=15d", regulators=["UK AISI", "US AISI", "EU AI Office"]), +] + +hubComponents = [ + hub_component("HUB-01", "UI", "React + Next.js + GraphQL Apollo Client", tech=["React 19", "Next.js 15", "Apollo Client"]), + hub_component("HUB-02", "API", "GraphQL Federation gateway", tech=["Apollo Gateway", "Hasura", "GraphQL-Mesh"]), + hub_component("HUB-03", "API", "REST adapter for legacy GRC tools", tech=["OpenAPI 3.1"]), + hub_component("HUB-04", "Domain", "Model Inventory service (Go)", patterns=["DDD", "Event sourcing"]), + hub_component("HUB-05", "Domain", "MRM Workbench service (Java/Spring Boot)", patterns=["CQRS", "Saga"]), + hub_component("HUB-06", "Domain", "Risk Register service (Go)", patterns=["Event sourcing"]), + hub_component("HUB-07", "Domain", "Policy Catalog service (Go) + OPA integration", patterns=["GitOps"]), + hub_component("HUB-08", "Domain", "Evidence Pack service (Python)", outputs=["PDF", "JSON-LD", "C2PA-signed bundle"]), + hub_component("HUB-09", "Domain", "Decision Log Explorer (Trino + Iceberg)", tech=["Trino", "Iceberg", "ksqlDB"]), + hub_component("HUB-10", "Domain", "AGI Watchtower service (Python/Rust)", outputs=["Capability dashboards", "Threshold alerts"]), + hub_component("HUB-11", "Domain", "Red-Team Findings service (Go)", integrations=["Jira", "ServiceNow"]), + hub_component("HUB-12", "Domain", "Regulator Portal service (Go)", auth=["OIDC", "mTLS", "WebAuthn"]), + hub_component("HUB-13", "Event Bus", "Apache Kafka with Schema Registry", tech=["Kafka 3.7+", "Confluent SR", "tiered storage"]), + hub_component("HUB-14", "Data Plane", "PostgreSQL + Iceberg on S3/GCS/Azure", tech=["PostgreSQL 16", "Apache Iceberg"]), + hub_component("HUB-15", "Identity", "Keycloak OIDC + OPA authorization", tech=["Keycloak 25+", "OPA", "OPAL"]), + hub_component("HUB-16", "Observability", "OpenTelemetry + Prometheus + Grafana + Splunk/Datadog", tech=["OTel", "Prometheus", "Grafana"]), +] + +# ========================================================================= +# Tail: schemas, code, KPIs, RCM, traceability, dataFlows, regulators, +# privacy, deployment, rollout90, roadmap, evidencePack, exec summary +# ========================================================================= +schemas = [ + {"sid": "SCH-01", "name": "AIGovDecisionEvent", "fields": ["decisionId","modelId","tier","userId(tok)","timestamp","inputHash","outputHash","explanationRef","consentId","purposeId","piiClass","fairnessFlag","approverIds","opaBundleHash"]}, + {"sid": "SCH-02", "name": "ModelInventoryRecord", "fields": ["modelId","name","version","tier","owner","lob","useCase","euAiActClass","mrmStatus","piiHandling","createdAt","retiredAt"]}, + {"sid": "SCH-03", "name": "MRMValidationReport", "fields": ["reportId","modelId","validator","conceptualSoundness","ongoingMonitoring","outcomesAnalysis","fairnessReport","approvalStatus","approverIds","date"]}, + {"sid": "SCH-04", "name": "RiskRegisterEntry", "fields": ["riskId","category","description","likelihood","impact","inherent","controls","residual","owner","reviewDate"]}, + {"sid": "SCH-05", "name": "PolicyDoc", "fields": ["pid","domain","statement","owner","cadence","evidence","version","effectiveDate","supersedes"]}, + {"sid": "SCH-06", "name": "EvidencePack", "fields": ["epid","regulator","period","artifacts[]","hash","signedBy","format"]}, + {"sid": "SCH-07", "name": "RedTeamFinding", "fields": ["findingId","modelId","vector","technique","severity","cvss","exploitability","impact","remediationPlan","sla","status"]}, + {"sid": "SCH-08", "name": "ContainmentEvent", "fields": ["eventId","tier","trigger","action","approvers[]","kineticInvoked","aisiNotified","timestamp"]}, + {"sid": "SCH-09", "name": "OPAPolicyBundle", "fields": ["bundleId","sha256","mlDsaSig","sourceRepo","sourceCommit","deployedAt","version"]}, + {"sid": "SCH-10", "name": "KafkaTopicSpec", "fields": ["tid","name","schema","retention","partitions","replication","minISR","acks"]}, + {"sid": "SCH-11", "name": "DriftAlert", "fields": ["alertId","modelId","metric","value","threshold","window","severity","action"]}, + {"sid": "SCH-12", "name": "FairnessMetric", "fields": ["metricId","modelId","protectedClass","metric","value","threshold","timestamp"]}, + {"sid": "SCH-13", "name": "ConsentEvent", "fields": ["consentId","customerId(tok)","purpose","status","timestamp","jurisdictions[]"]}, + {"sid": "SCH-14", "name": "DPIA", "fields": ["dpiaId","modelId","dataClasses","processing","necessity","proportionality","risks","mitigations","approvedBy","date"]}, + {"sid": "SCH-15", "name": "RegulatorNotification", "fields": ["notifId","regulator","category","severity","reportedAt","deadline","contentHash","ackRef"]}, + {"sid": "SCH-16", "name": "TrainingRun", "fields": ["runId","modelId","datasetIds[]","flops","tokens","start","end","seed","artifacts[]","aisiNotified"]}, +] + +code = [ + {"cid": "CODE-01", "lang": "rego", "name": "policies/admission/require_signed_image.rego", "purpose": "Cosign signature admission gate"}, + {"cid": "CODE-02", "lang": "rego", "name": "policies/deployment/mrm_validation_gate.rego", "purpose": "MRM validation status gate"}, + {"cid": "CODE-03", "lang": "rego", "name": "policies/runtime/data_purpose_limitation.rego", "purpose": "GDPR purpose limitation check"}, + {"cid": "CODE-04", "lang": "rego", "name": "policies/agi/quorum_3of5.rego", "purpose": "Frontier 3-of-5 quorum"}, + {"cid": "CODE-05", "lang": "yaml", "name": "kyverno/require-cosign.yaml", "purpose": "Kyverno Cosign verify policy"}, + {"cid": "CODE-06", "lang": "yaml", "name": "cilium/default-deny.yaml", "purpose": "Cilium default-deny NetworkPolicy"}, + {"cid": "CODE-07", "lang": "yaml", "name": "falco/rules-ai.yaml", "purpose": "Falco rules for AI workload anomalies"}, + {"cid": "CODE-08", "lang": "python", "name": "drift/psi_monitor.py", "purpose": "PSI/KS drift monitor producing aigov.drift-alerts"}, + {"cid": "CODE-09", "lang": "python", "name": "fairness/disparate_impact.py", "purpose": "Quarterly disparate-impact analysis"}, + {"cid": "CODE-10", "lang": "python", "name": "redteam/prompt_injection.py", "purpose": "Prompt injection harness with OWASP LLM01 vectors"}, + {"cid": "CODE-11", "lang": "go", "name": "services/decisionlog/main.go", "purpose": "Decision log producer to aigov.decisions"}, + {"cid": "CODE-12", "lang": "go", "name": "services/worm-sealer/main.go", "purpose": "WORM sealer with ML-DSA + merkle"}, + {"cid": "CODE-13", "lang": "tla+", "name": "specs/control_plane.tla", "purpose": "TLA+ spec for AGI control plane invariants"}, + {"cid": "CODE-14", "lang": "graphql", "name": "schema/hub.graphql", "purpose": "Federated GraphQL schema for Hub"}, + {"cid": "CODE-15", "lang": "yaml", "name": "argo-cd/aigov-app.yaml", "purpose": "Argo CD GitOps app for Hub"}, +] + +kpis = [ + {"kid": "KPI-01", "name": "AIMS-Coverage", "target": ">=0.95", "cadence": "Monthly"}, + {"kid": "KPI-02", "name": "MRGI", "target": ">=0.95", "cadence": "Monthly"}, + {"kid": "KPI-03", "name": "DRI", "target": ">=0.95", "cadence": "Per decision"}, + {"kid": "KPI-04", "name": "CCS", "target": ">=0.95", "cadence": "Monthly"}, + {"kid": "KPI-05", "name": "ARI", "target": ">=0.9 frontier", "cadence": "Weekly"}, + {"kid": "KPI-06", "name": "CSI", "target": ">=0.95 T3/T4", "cadence": "Continuous"}, + {"kid": "KPI-07", "name": "RTRI", "target": ">=0.9", "cadence": "Per red-team cycle"}, + {"kid": "KPI-08", "name": "CDC-Score", "target": ">=0.9", "cadence": "Quarterly"}, + {"kid": "KPI-09", "name": "RCI", "target": "=1.0", "cadence": "Per regulator engagement"}, + {"kid": "KPI-10", "name": "Models registered in Hub", "target": "100%", "cadence": "Monthly"}, + {"kid": "KPI-11", "name": "T2+ models with red-team report", "target": "100%", "cadence": "Monthly"}, + {"kid": "KPI-12", "name": "DPIAs current (T2+ PII)", "target": "100%", "cadence": "Monthly"}, + {"kid": "KPI-13", "name": "MRM validations on time", "target": ">=98%", "cadence": "Monthly"}, + {"kid": "KPI-14", "name": "Kafka audit log durability", "target": "=11x9s", "cadence": "Continuous"}, + {"kid": "KPI-15", "name": "WORM seal verification pass", "target": "100%", "cadence": "Daily"}, + {"kid": "KPI-16", "name": "OPA decision latency p99", "target": "<=5ms", "cadence": "Continuous"}, + {"kid": "KPI-17", "name": "K8s admission deny rate (false-positive)", "target": "<=1%", "cadence": "Monthly"}, + {"kid": "KPI-18", "name": "Critical red-team SLA compliance", "target": ">=95% <=7d", "cadence": "Monthly"}, + {"kid": "KPI-19", "name": "Frontier capability threshold breaches", "target": "0 unreported", "cadence": "Continuous"}, + {"kid": "KPI-20", "name": "Kinetic override drills completed", "target": ">=4/y", "cadence": "Quarterly"}, + {"kid": "KPI-21", "name": "AISI notifications on time", "target": "100% <=24h", "cadence": "Per event"}, + {"kid": "KPI-22", "name": "EU AI Office notifications on time", "target": "100% <=15d", "cadence": "Per event"}, + {"kid": "KPI-23", "name": "SEC 8-K materiality decisions on time", "target": "100% <=4 BD", "cadence": "Per event"}, + {"kid": "KPI-24", "name": "DORA major incident reports on time", "target": "100% <=4h", "cadence": "Per event"}, + {"kid": "KPI-25", "name": "Consumer Duty foreseeable-harm assessments", "target": "100%", "cadence": "Annual"}, + {"kid": "KPI-26", "name": "Disparate-impact quarterly tests", "target": "100% credit/HR", "cadence": "Quarterly"}, + {"kid": "KPI-27", "name": "FCRA adverse-action notices <=30d", "target": "100%", "cadence": "Per event"}, + {"kid": "KPI-28", "name": "PQC migration coverage", "target": ">=80% by 2028, 100% by 2030", "cadence": "Annual"}, + {"kid": "KPI-29", "name": "ISO 42001 surveillance audits", "target": "no major NCRs", "cadence": "Annual"}, + {"kid": "KPI-30", "name": "Board AI Risk Committee meetings", "target": ">=4/y", "cadence": "Quarterly"}, +] + +riskControlMatrix = [ + {"rid": "R-01", "risk": "Unauthorized AGI capability emergence", "likelihood": "Low", "impact": "Catastrophic", "control": "T4 quorum + kinetic + AISI", "owner": "Board AI Risk Cmt"}, + {"rid": "R-02", "risk": "Model risk capital misstatement", "likelihood": "Med", "impact": "High", "control": "SR 11-7 + ICAAP", "owner": "CRO"}, + {"rid": "R-03", "risk": "GDPR Art-22 violation", "likelihood": "Med", "impact": "High", "control": "DPIA + Art-22 path", "owner": "DPO"}, + {"rid": "R-04", "risk": "FCRA/ECOA disparate impact", "likelihood": "Med", "impact": "High", "control": "Quarterly DI tests", "owner": "CCO"}, + {"rid": "R-05", "risk": "EU AI Act high-risk non-compliance", "likelihood": "Med", "impact": "High", "control": "Art. 9/10/14/15 controls", "owner": "CCO"}, + {"rid": "R-06", "risk": "FCA Consumer Duty breach", "likelihood": "Med", "impact": "High", "control": "Foreseeable-harm assess", "owner": "SMF-AI"}, + {"rid": "R-07", "risk": "Kafka audit tampering", "likelihood": "Low", "impact": "High", "control": "WORM + PQC seal + indep verifier", "owner": "CISO"}, + {"rid": "R-08", "risk": "K8s container escape", "likelihood": "Low", "impact": "High", "control": "PSA restricted + Falco + Tetragon", "owner": "CISO"}, + {"rid": "R-09", "risk": "OPA policy bypass", "likelihood": "Low", "impact": "High", "control": "Signed bundles + GitOps + decision log", "owner": "CISO"}, + {"rid": "R-10", "risk": "Prompt injection causing data leak", "likelihood": "High", "impact": "Med", "control": "Red-team RT-01..03 + OPA runtime", "owner": "CDAO"}, + {"rid": "R-11", "risk": "Training data poisoning", "likelihood": "Low", "impact": "High", "control": "Data provenance + RT-09", "owner": "CDAO"}, + {"rid": "R-12", "risk": "DORA major incident missed deadline", "likelihood": "Low", "impact": "High", "control": "IR runbook + DORA <=4h SLA", "owner": "CISO"}, + {"rid": "R-13", "risk": "SEC cyber disclosure miss", "likelihood": "Low", "impact": "High", "control": "Materiality playbook <=4 BD", "owner": "CFO+CCO"}, + {"rid": "R-14", "risk": "Third-party AI vendor failure", "likelihood": "Med", "impact": "Med", "control": "Critical TPRM register per DORA", "owner": "Head TPRM"}, + {"rid": "R-15", "risk": "PQC migration delay", "likelihood": "Med", "impact": "Med", "control": "Hybrid TLS + roadmap per CNSA 2.0", "owner": "CISO"}, + {"rid": "R-16", "risk": "MAS/HKMA fairness non-compliance APAC", "likelihood": "Med", "impact": "Med", "control": "FEAT + GP-1/GS-2 controls", "owner": "Regional CCO APAC"}, +] + +traceability = [ + {"tid": "T-01", "control": "AIMS Policy", "regime": "ISO 42001", "clause": "5.2", "evidence": "Board-signed AI Policy"}, + {"tid": "T-02", "control": "Risk Mgmt", "regime": "NIST AI RMF", "clause": "MAP-2.1", "evidence": "AI Risk Register"}, + {"tid": "T-03", "control": "EU AI Act Art. 9", "regime": "EU AI Act", "clause": "Art. 9", "evidence": "Risk mgmt system docs"}, + {"tid": "T-04", "control": "EU AI Act Art. 10", "regime": "EU AI Act", "clause": "Art. 10", "evidence": "Data governance docs"}, + {"tid": "T-05", "control": "EU AI Act Art. 14", "regime": "EU AI Act", "clause": "Art. 14", "evidence": "Human oversight runbook"}, + {"tid": "T-06", "control": "EU AI Act Art. 15", "regime": "EU AI Act", "clause": "Art. 15", "evidence": "Accuracy/robustness/cyber report"}, + {"tid": "T-07", "control": "GPAI Art. 53 tech doc", "regime": "EU AI Act", "clause": "Art. 53", "evidence": "GPAI tech doc + copyright policy"}, + {"tid": "T-08", "control": "GPAI Art. 55 systemic", "regime": "EU AI Act", "clause": "Art. 55", "evidence": "Frontier evals + incident reports"}, + {"tid": "T-09", "control": "GDPR DPIA", "regime": "GDPR", "clause": "Art. 35", "evidence": "DPIA registry"}, + {"tid": "T-10", "control": "GDPR Art-22", "regime": "GDPR", "clause": "Art. 22", "evidence": "Art-22 invocation logs"}, + {"tid": "T-11", "control": "FCRA adverse action", "regime": "FCRA", "clause": "615(a)", "evidence": "Notice generation logs"}, + {"tid": "T-12", "control": "ECOA Reg-B", "regime": "ECOA", "clause": "1002.9", "evidence": "Disparate-impact report"}, + {"tid": "T-13", "control": "SR 11-7 effective challenge", "regime": "US Fed", "clause": "SR 11-7 V", "evidence": "Independent validation"}, + {"tid": "T-14", "control": "OCC 2011-12", "regime": "OCC", "clause": "2011-12.III", "evidence": "Model dev doc"}, + {"tid": "T-15", "control": "FCA Consumer Duty", "regime": "FCA", "clause": "PRIN 2A", "evidence": "Consumer Duty Board report"}, + {"tid": "T-16", "control": "SMCR SMF-AI", "regime": "FCA/PRA", "clause": "SMF-AI", "evidence": "SMF-AI statement of responsibilities"}, + {"tid": "T-17", "control": "MAS FEAT", "regime": "MAS", "clause": "FEAT", "evidence": "FEAT principles attestation"}, + {"tid": "T-18", "control": "HKMA GP-1/GS-2", "regime": "HKMA", "clause": "GP-1+GS-2", "evidence": "AI governance attestation"}, + {"tid": "T-19", "control": "DORA major incident", "regime": "EU DORA", "clause": "Art. 19", "evidence": "Incident reporting log"}, + {"tid": "T-20", "control": "SEC 17a-4 WORM", "regime": "SEC", "clause": "17 CFR 240.17a-4(f)", "evidence": "WORM attestation"}, +] + +dataFlows = [ + {"fid": "DF-01", "src": "Feature store", "sink": "Model inference", "class": "PII tokenized", "purpose": "decisioning"}, + {"fid": "DF-02", "src": "Model inference", "sink": "Kafka aigov.decisions", "class": "tokenized", "purpose": "audit"}, + {"fid": "DF-03", "src": "Kafka aigov.decisions", "sink": "WORM S3 Object Lock", "class": "sealed", "purpose": "retention"}, + {"fid": "DF-04", "src": "Kafka aigov.decisions", "sink": "Trino on Iceberg", "class": "tokenized", "purpose": "query"}, + {"fid": "DF-05", "src": "Trino", "sink": "Hub Decision Log Explorer", "class": "RBAC-filtered", "purpose": "UI"}, + {"fid": "DF-06", "src": "Hub", "sink": "Regulator Portal", "class": "read-only scoped", "purpose": "regulator"}, + {"fid": "DF-07", "src": "GitHub policies repo", "sink": "OPAL distribution", "class": "signed", "purpose": "policy"}, + {"fid": "DF-08", "src": "OPAL", "sink": "OPA sidecars + Gatekeeper", "class": "signed bundle", "purpose": "enforce"}, + {"fid": "DF-09", "src": "OPA", "sink": "Kafka aigov.access + aigov.policy-changes", "class": "decision log", "purpose": "audit"}, + {"fid": "DF-10", "src": "MRM Workbench", "sink": "Hub Model Inventory", "class": "metadata", "purpose": "lifecycle"}, + {"fid": "DF-11", "src": "Red-team tools", "sink": "Kafka aigov.red-team-findings", "class": "findings", "purpose": "remediation"}, + {"fid": "DF-12", "src": "AGI Watchtower evals", "sink": "Kafka aigov.eval-results + Hub", "class": "capability scores", "purpose": "containment"}, +] + +regulators = [ + {"reg": "EU AI Office", "scope": "EU AI Act + GPAI", "cadence": "Quarterly + on incident"}, + {"reg": "European Data Protection Board", "scope": "GDPR", "cadence": "On incident + on request"}, + {"reg": "FCA", "scope": "Consumer Duty + SMCR", "cadence": "Annual + SS1/23"}, + {"reg": "PRA", "scope": "SS1/23 model risk", "cadence": "Annual"}, + {"reg": "Bank of England", "scope": "Systemic + DORA-equivalent", "cadence": "Annual"}, + {"reg": "ECB SSM", "scope": "Eurozone banking", "cadence": "Annual SREP"}, + {"reg": "US Federal Reserve", "scope": "SR 11-7", "cadence": "Annual + supervisory cycle"}, + {"reg": "OCC", "scope": "OCC 2011-12", "cadence": "Annual"}, + {"reg": "FDIC", "scope": "US insured banks", "cadence": "Annual"}, + {"reg": "CFPB", "scope": "FCRA/ECOA consumer", "cadence": "On complaint + sweeps"}, + {"reg": "SEC", "scope": "17a-4 + 10-K/8-K + cyber", "cadence": "Per event + annual"}, + {"reg": "FINRA", "scope": "3110/4511", "cadence": "Annual exam"}, + {"reg": "MAS", "scope": "FEAT + TRM", "cadence": "Annual"}, + {"reg": "HKMA", "scope": "GP-1 + GS-2", "cadence": "Annual"}, + {"reg": "OSFI", "scope": "E-23", "cadence": "Annual"}, + {"reg": "FINMA", "scope": "AI guidance", "cadence": "Annual"}, +] + +privacy = { + "dpiaPolicy": "Required for all T2+ models with PII or special category", + "rightsOps": ["Access (Art. 15)", "Rectification (Art. 16)", "Erasure (Art. 17 / right to be forgotten)", "Restriction (Art. 18)", "Portability (Art. 20)", "Object (Art. 21)", "Art-22 human review"], + "transferMechanisms": ["EU SCC 2021/914", "UK IDTA", "Adequacy decisions", "BCRs"], + "minimization": "Purpose-limitation enforced via OPA-04..09; data minimization audited annually", +} + +deployment = { + "tiering": ["T0 sandbox -> T1 staging -> T2 canary <=1% -> T3 prod Nitro Enclaves -> T4 frontier air-gapped"], + "gitops": "Argo CD + Crossplane + Terraform; signed manifests; environment promotion via PR", + "regions": ["us-east-1", "us-west-2", "eu-west-1", "eu-central-1", "ap-southeast-1", "ap-northeast-1", "uk-south", "ca-central-1"], + "dr": {"rto": "<=4h Hub UI; <=1h decision log", "rpo": "<=15min", "drills": "quarterly full failover + tabletop"}, +} + +rollout90 = [ + {"day": "0-30", "focus": "Foundation", "deliverables": ["AI Policy Board-signed", "AI Risk Register v1", "AI Governance Hub MVP", "ISO 42001 gap assessment", "Model inventory bootstrapped", "Kafka audit topics aigov.decisions + policy-changes live"]}, + {"day": "31-60", "focus": "Controls", "deliverables": ["OPA admission gates in dev/staging", "MRM Workbench T1 models loaded", "WORM tier deployed in 1 region", "DPIA registry populated", "Red-team baseline run on top-10 T2 models", "FCA Consumer Duty foreseeable-harm framework"]}, + {"day": "61-90", "focus": "Production + Regulator", "deliverables": ["OPA gates in prod for T2+", "WORM multi-region", "First quarterly Board AI Risk Cmt", "FCRA/ECOA disparate-impact pipeline live", "Regulator portal live (read-only)", "First evidence pack generated"]}, +] + +roadmap = [ + {"yr": "2026", "milestone": "Hub GA; ISO 42001 stage-1 audit; OPA enterprise rollout; first GPAI Art. 55 attestation"}, + {"yr": "2027", "milestone": "ISO 42001 certified; full EU AI Act high-risk coverage; PQC ML-DSA on all seals; FCA Consumer Duty embedded"}, + {"yr": "2028", "milestone": "Full Kafka+WORM regulator portals; T4 frontier evals operationalized; AISI MoUs active; PQC >=80%"}, + {"yr": "2029", "milestone": "Federated PETs + confidential containers default for T3; cross-border data residency 100% OPA-enforced"}, + {"yr": "2030", "milestone": "PQC 100% across all sealing + TLS; AGI containment T4 industrialized; Hub federation across G-SIFI peers"}, +] + +evidencePack = [ + {"epid": "EP-01", "name": "AIMS Manual + Scope Statement", "format": "PDF + JSON-LD"}, + {"epid": "EP-02", "name": "AI Risk Register snapshot", "format": "CSV + signed"}, + {"epid": "EP-03", "name": "Model Inventory snapshot", "format": "CSV + JSON"}, + {"epid": "EP-04", "name": "MRM Validation Reports (period)", "format": "PDF bundle"}, + {"epid": "EP-05", "name": "DPIA Registry snapshot", "format": "CSV + JSON"}, + {"epid": "EP-06", "name": "Fairness/Disparate-Impact Reports", "format": "PDF"}, + {"epid": "EP-07", "name": "Red-Team Findings + Remediation", "format": "PDF + JSON"}, + {"epid": "EP-08", "name": "Kafka WORM Seal Verifications", "format": "JSON-LD signed"}, + {"epid": "EP-09", "name": "OPA Decision Log extracts", "format": "Parquet + signed manifest"}, + {"epid": "EP-10", "name": "Containment Event Log + AISI Notifications", "format": "JSON-LD signed"}, + {"epid": "EP-11", "name": "GPAI Art. 53 Technical Documentation", "format": "PDF + JSON-LD"}, + {"epid": "EP-12", "name": "GPAI Art. 55 Evals + Incident Reports", "format": "PDF + JSON-LD"}, + {"epid": "EP-13", "name": "FCRA/ECOA Adverse Action Notice Logs", "format": "Parquet"}, + {"epid": "EP-14", "name": "Consumer Duty Board Report", "format": "PDF"}, + {"epid": "EP-15", "name": "ICAAP AI Risk Section", "format": "PDF"}, + {"epid": "EP-16", "name": "PQC Migration Status Report", "format": "PDF + JSON"}, +] + +executiveSummary = { + "thesis": "Enterprise AI/AGI governance for Fortune 500 / Global 2000 / G-SIFIs requires an integrated operating model anchored on ISO/IEC 42001 AIMS, mapped bidirectionally to NIST AI RMF + EU AI Act + GDPR + sectoral financial regimes (SR 11-7, Basel, FCA Consumer Duty, MAS FEAT, HKMA GP-1/GS-2), operationalized via Kafka audit + WORM + PQC + Kubernetes security + OPA/Rego + MRM platform + red-teaming + AGI containment T0-T4, surfaced through a single AI Governance Hub.", + "investment": "USD 180-500M / 5y; NPV USD 500-1500M risk-adjusted", + "headlineRisks": ["Frontier AGI capability emergence", "EU AI Act high-risk non-compliance", "FCRA/ECOA disparate impact", "Kafka audit tampering", "PQC migration delay"], + "ninetyDay": ["Board-signed AI Policy + RAS", "Hub MVP live", "ISO 42001 gap assessment", "OPA admission gates in prod", "First regulator evidence pack"], +} + +# Final assembly +DOC["modules"] = MODULES +DOC["policies"] = policies +DOC["controls"] = controls +DOC["kafkaTopics"] = kafkaTopics +DOC["k8sControls"] = k8sControls +DOC["opaPolicies"] = opaPolicies +DOC["wormControls"] = wormControls +DOC["mrmArtifacts"] = mrmArtifacts +DOC["redTeams"] = redTeams +DOC["agiContainments"] = agiContainments +DOC["hubComponents"] = hubComponents +DOC["schemas"] = schemas +DOC["code"] = code +DOC["kpis"] = kpis +DOC["riskControlMatrix"] = riskControlMatrix +DOC["traceability"] = traceability +DOC["dataFlows"] = dataFlows +DOC["regulators"] = regulators +DOC["privacy"] = privacy +DOC["deployment"] = deployment +DOC["rollout90"] = rollout90 +DOC["roadmap"] = roadmap +DOC["evidencePack"] = evidencePack +DOC["executiveSummary"] = executiveSummary + +DOC["counts"] = { + "modules": len(MODULES), + "sections": sum(len(m["sections"]) for m in MODULES), + "policies": len(policies), + "controls": len(controls), + "kafkaTopics": len(kafkaTopics), + "k8sControls": len(k8sControls), + "opaPolicies": len(opaPolicies), + "wormControls": len(wormControls), + "mrmArtifacts": len(mrmArtifacts), + "redTeams": len(redTeams), + "agiContainments": len(agiContainments), + "hubComponents": len(hubComponents), + "schemas": len(schemas), + "code": len(code), + "kpis": len(kpis), + "riskControlMatrix": len(riskControlMatrix), + "traceability": len(traceability), + "dataFlows": len(dataFlows), + "regulators": len(regulators), + "rollout90": len(rollout90), + "roadmap": len(roadmap), + "evidencePack": len(evidencePack), +} + +os.makedirs(os.path.dirname(OUT), exist_ok=True) +with open(OUT, "w") as f: + json.dump(DOC, f, indent=2) + +print(f"WP-058 JSON written: {OUT}") +print(f"Size: {os.path.getsize(OUT):,} bytes ({os.path.getsize(OUT)/1024:.1f} KB)") +print(f"Counts: {DOC['counts']}") diff --git a/rag-agentic-dashboard/public/enterprise-aigov-framework.html b/rag-agentic-dashboard/public/enterprise-aigov-framework.html new file mode 100644 index 0000000..cc92e83 --- /dev/null +++ b/rag-agentic-dashboard/public/enterprise-aigov-framework.html @@ -0,0 +1,132 @@ + + +Enterprise AI/AGI Governance Framework for Large Financial & Fortune 500 Enterprises (2026-2030) + + +
    +

    Enterprise AI/AGI Governance Framework for Large Financial & Fortune 500 Enterprises (2026-2030)

    +
    docRef ENTERPRISE-AIGOV-FRAMEWORK-WP-058 · v1.0.0 · regulator-submission-grade · Confidential / Restricted — Board, CRO, CCO, CISO, CDAO, Group Internal Audit, External Regulators (on request)
    +
    Horizon: 2026-2030 · API prefix: /api/enterprise-aigov-framework · builds on WP-035 · WP-040 · WP-045 · WP-050 · WP-054 · WP-055 · WP-056 · WP-057
    +
    +9 modules45 sections15 policies25 controls12 kafkaTopics15 k8sControls15 opaPolicies12 wormControls15 mrmArtifacts15 redTeams15 agiContainments16 hubComponents16 schemas15 code30 kpis16 riskControlMatrix20 traceability12 dataFlows16 regulators3 rollout905 roadmap16 evidencePack +
    +
    +
    + +
    + +

    Executive Summary

    +

    Thesis: Enterprise AI/AGI governance for Fortune 500 / Global 2000 / G-SIFIs requires an integrated operating model anchored on ISO/IEC 42001 AIMS, mapped bidirectionally to NIST AI RMF + EU AI Act + GDPR + sectoral financial regimes (SR 11-7, Basel, FCA Consumer Duty, MAS FEAT, HKMA GP-1/GS-2), operationalized via Kafka audit + WORM + PQC + Kubernetes security + OPA/Rego + MRM platform + red-teaming + AGI containment T0-T4, surfaced through a single AI Governance Hub.

    +

    Investment: USD 180-500M / 5y; NPV USD 500-1500M risk-adjusted

    +

    Headline risks: Frontier AGI capability emergence, EU AI Act high-risk non-compliance, FCRA/ECOA disparate impact, Kafka audit tampering, PQC migration delay

    +

    First 90 days: Board-signed AI Policy + RAS, Hub MVP live, ISO 42001 gap assessment, OPA admission gates in prod, First regulator evidence pack

    +
    + + +

    Strategic Directive

    +

    Scope: End-to-end enterprise AI/AGI governance operating model for Fortune 500 / Global 2000 / G-SIFIs spanning policy, control, risk, compliance, security, model risk, third-party, AGI containment, and AI Governance Hub architecture

    +
    Outcomes
    • ISO/IEC 42001:2023 certified AIMS by 2027 across all material AI systems
    • NIST AI RMF 1.0 + AI 600-1 generative profile mapped to >=95% of in-scope models
    • EU AI Act 2026 Article 6/9/10/14/15 + GPAI 53/55 obligations fully evidenced
    • GDPR DPIA + Article 22 + FCRA/ECOA adverse-action automation in production
    • Basel III/IV + SR 11-7 + OCC 2011-12 model risk lifecycle managed in single MRM platform
    • FCA Consumer Duty + SMCR SMF-attested AI accountability operating model
    • MAS FEAT + HKMA GP-1/GS-2 fairness, ethics, accountability, transparency in production
    • Kafka audit log with WORM + PQC sealing across all model decisions
    • Kubernetes + container security with policy-as-code (OPA/Rego) at admission and runtime
    • AGI containment T0-T4 with 3-of-5 quorum + kinetic override and AI Safety Institute notifications
    +
    Do NOT
    • Do NOT deploy any AI/AGI capability without Enterprise AI Governance Hub registration, ISO 42001 risk assessment, and model risk tier classification
    • Do NOT bypass Kafka audit logging, OPA/Rego policy gates, WORM/PQC sealing, or 3-of-5 frontier quorum
    +
    + +

    Regulatory Regimes (28)

    • ISO/IEC 42001:2023 AIMS
    • ISO/IEC 23894:2023 AI Risk
    • ISO/IEC 27001:2022 ISMS
    • ISO/IEC 27701:2019 PIMS
    • NIST AI RMF 1.0
    • NIST AI 600-1 Generative AI Profile
    • NIST SP 800-53 Rev.5
    • NIST SP 800-218 SSDF
    • OECD AI Principles 2019/2024
    • EU AI Act 2024/1689
    • EU AI Act GPAI Art. 53/55
    • EU GDPR + Art. 22
    • EU DORA
    • EU NIS2
    • EU CRA
    • US FCRA + ECOA Reg-B
    • US Fed SR 11-7 + OCC 2011-12
    • Basel III/IV + ICAAP
    • US SEC 17a-4 + 10-K/8-K
    • FINRA 3110/4511
    • UK FCA Consumer Duty
    • UK SMCR + SS1/23
    • MAS FEAT + TRM 2021
    • HKMA GP-1 + GS-2
    • OSFI E-23
    • FINMA AI Guidance
    • G7 Hiroshima Process
    • Bletchley/Seoul/Paris Declarations
    + +

    Performance Indices

    • AIMS-Coverage: >=0.95 (ISO 42001 controls)
    • MRGI: >=0.95 (Model Risk Governance Index, SR 11-7 + OCC 2011-12)
    • DRI: >=0.95 (Decision Reproducibility, n=10)
    • CCS: >=0.95 (Control Coverage Score)
    • ARI: >=0.9 (Alignment Robustness Index, frontier)
    • CSI: >=0.95 (Containment Sufficiency, T3/T4)
    • RTRI: >=0.9 (Red-Team Resilience Index)
    • CDC-Score: >=0.9 (FCA Consumer Duty compliance)
    • RCI: =1.0 (Regulator Confidence Index)
    + +

    Tiers (T0-T4)

    • T0: Sandbox - isolated VPC, synthetic data only, no production traffic
    • T1: Staging - shadow mode, real data, no customer impact
    • T2: Canary - <=1% production traffic, automated rollback
    • T3: Production - Nitro Enclaves + KMS + dual control + full audit
    • T4: Frontier Air-Gapped - 3-of-5 quorum + kinetic override + AI Safety Institute notification
    + +

    Severity Levels

    • SEV-0: Civilizational/systemic - EU AI Office notification <=15d, AISI notification <=24h
    • SEV-1: Major - SEC 8-K <=4 BD, DORA <=4h, FCA <=72h
    • SEV-2: Material - regulator notification <=72h
    • SEV-3: Operational - internal escalation <=10 BD
    + +

    Investment Envelope

    +

    Envelope: USD 180-500M / 5y (Fortune 500 / G-SIFI tier) · NPV: USD 500-1500M (5y, risk-adjusted)

    +
    Drivers
    • MRM platform consolidation
    • AI Governance Hub build
    • Kafka audit + WORM/PQC sealing
    • Kubernetes + OPA/Rego at scale
    • AGI containment T3/T4 enclaves
    • Red-teaming program
    • Regulator attestation tooling
    +
    + +

    Privacy & Data Protection

    +
    dpiaPolicy: Required for all T2+ models with PII or special category
    rightsOps
    • Access (Art. 15)
    • Rectification (Art. 16)
    • Erasure (Art. 17 / right to be forgotten)
    • Restriction (Art. 18)
    • Portability (Art. 20)
    • Object (Art. 21)
    • Art-22 human review
    transferMechanisms
    • EU SCC 2021/914
    • UK IDTA
    • Adequacy decisions
    • BCRs
    minimization: Purpose-limitation enforced via OPA-04..09; data minimization audited annually
    +
    + +

    Deployment Model

    +
    tiering
    • T0 sandbox -> T1 staging -> T2 canary <=1% -> T3 prod Nitro Enclaves -> T4 frontier air-gapped
    gitops: Argo CD + Crossplane + Terraform; signed manifests; environment promotion via PR
    regions
    • us-east-1
    • us-west-2
    • eu-west-1
    • eu-central-1
    • ap-southeast-1
    • ap-northeast-1
    • uk-south
    • ca-central-1
    dr
    • rto: <=4h Hub UI; <=1h decision log
    • rpo: <=15min
    • drills: quarterly full failover + tabletop
    +
    + +

    M1 — ISO/IEC 42001 AIMS + NIST AI RMF + OECD + EU AI Act Foundation

    Integrated AI Management System anchored on ISO/IEC 42001:2023 with NIST AI RMF 1.0 functions (Govern/Map/Measure/Manage), OECD AI Principles, and EU AI Act 2024/1689 Article 6/9/10/14/15 + GPAI 53/55 mappings.

    M1.1. ISO/IEC 42001 AIMS Architecture

    policyCommit: Board-attested AI Policy, AI Objectives, AI Risk Appetite Statement signed annually by Group CEO + Group CRO
    structure: AIMS clauses 4-10 mapped to enterprise functions: Context (Strategy), Leadership (CDAO+CRO), Planning (PMO), Support (HR+Procurement+IT), Operation (Lines of Business), Performance (Internal Audit), Improvement (CCO)

    M1.2. NIST AI RMF 1.0 + AI 600-1 Generative Profile

    functions
    • GOVERN
    • MAP
    • MEASURE
    • MANAGE
    generativeProfile: NIST AI 600-1 mapped to all FM/LLM use cases with synthetic content provenance (C2PA)
    crossWalk: Bidirectional crosswalk ISO 42001 <-> NIST AI RMF <-> EU AI Act maintained in MRM platform

    M1.3. OECD AI Principles 2019/2024 Implementation

    principles
    • Inclusive growth
    • Human-centered values
    • Transparency
    • Robustness
    • Accountability
    implementation: OECD-aligned model cards + system cards published for all T2+ systems with public-facing summary

    M1.4. EU AI Act 2024/1689 Compliance Layer

    riskClasses
    • Unacceptable (Art. 5)
    • High-risk (Art. 6/Annex III)
    • Limited-risk (Art. 50)
    • Minimal-risk
    highRiskObligations
    • Art. 9 risk mgmt
    • Art. 10 data governance
    • Art. 14 human oversight
    • Art. 15 accuracy/robustness/cybersecurity
    gpai
    • Art. 53 technical documentation + copyright policy
    • Art. 55 systemic-risk: evaluations, adversarial testing, cybersecurity, incident reporting
    timeline: Prohibited practices Feb 2025; GPAI Aug 2025; high-risk Aug 2026

    M1.5. Board + Executive Accountability

    committees
    • Board AI Risk Committee (quarterly)
    • Executive AI Governance Committee (monthly)
    • AI Ethics Council (monthly)
    • Model Risk Committee (weekly)
    roles
    • AI-SMF (SMF-AI): FCA-attested senior manager
    • CDAO: Operating accountability
    • CRO: Risk appetite owner
    • CCO: Regulatory engagement
    • CISO: Cybersecurity + integrity
    • GIA: Independent assurance

    M2 — Financial-Services Model Risk Management (SR 11-7 / OCC 2011-12 / Basel III/IV / ICAAP)

    Three-lines-of-defense model risk operating model with SR 11-7 conceptual soundness, ongoing monitoring, outcomes analysis; OCC 2011-12 effective challenge; Basel III/IV IRB/IMA model validation; ICAAP integration with Pillar 2.

    M2.1. Model Risk Lifecycle

    stages
    • Identification
    • Development
    • Validation
    • Approval
    • Implementation
    • Monitoring
    • Retirement
    tiering: Tier-1 (regulatory capital, P&L, capital plan) / Tier-2 (material business) / Tier-3 (limited scope) / Tier-4 (research)
    cadence: Tier-1 annual validation; Tier-2 biennial; Tier-3 every 3y; ongoing monitoring monthly for all

    M2.2. SR 11-7 + OCC 2011-12 Effective Challenge

    conceptualSoundness: Independent review of theory, assumptions, design choices
    ongoingMonitoring
    • Backtesting
    • Benchmarking
    • Sensitivity analysis
    • Stress testing
    outcomesAnalysis: Champion/challenger + counterfactual analysis on production decisions

    M2.3. Basel III/IV IRB/IMA + FRTB

    validation: Independent validation per SR 15-19/SR 15-18; quantitative review every cycle
    capitalImpact: MRM platform feeds Pillar 2 model risk capital add-on into ICAAP

    M2.4. AI/ML-Specific MRM Extensions

    extensions
    • Concept drift + data drift monitoring (PSI, KS, KL, Wasserstein)
    • Fairness across protected classes (FCRA/ECOA aligned)
    • Explainability evidence (SHAP/LIME/integrated gradients) per decision
    • Adversarial robustness testing (PGD, BIM, NLP attacks)
    • Provenance of training data + lineage to feature store

    M2.5. ICAAP / Pillar 2 Integration

    integration: Aggregate model risk capital + AI-specific add-ons fed into ICAAP Pillar 2 alongside operational, reputational, strategic risk
    governance: MRC quarterly review of capital adequacy; Board-attested annual ICAAP includes AI risk section

    M3 — Data Protection & Consumer Fairness (GDPR / FCRA / ECOA / FCA Consumer Duty)

    Privacy-by-design with GDPR Article 22 (automated decisions), FCRA adverse-action notices, ECOA Reg-B disparate impact testing, FCA Consumer Duty cross-cutting rules + four outcomes; MAS FEAT + HKMA GP-1 fairness operationalization.

    M3.1. GDPR Article 22 + DPIA Operationalization

    dpia: DPIA required for all T2+ models processing personal data; reviewed by DPO + Group Legal
    article22
    • prohibition: No solely automated decisions producing legal/significant effects without explicit consent or necessity
    • rights: ['Human intervention', 'Express view', 'Contest decision']
    • logging: All Art-22 invocations logged to Kafka audit topic
    crossBorder: EU SCC + UK IDTA + adequacy assessments documented in ROPA

    M3.2. FCRA + ECOA Reg-B Adverse-Action Automation

    adverseAction: Automated FCRA 615(a) + ECOA Reg-B section 1002.9 notices generated within 30 days of adverse decision
    reasonCodes: Top-N reason codes derived from SHAP attribution, mapped to plain-English statements vetted by Compliance Legal
    disparateImpact: Quarterly disparate-impact analysis on credit, hiring, insurance models (race, sex, age, national origin)

    M3.3. FCA Consumer Duty + Cross-Cutting Rules

    outcomes
    • Products & services
    • Price & value
    • Consumer understanding
    • Consumer support
    crossCutting
    • Act in good faith
    • Avoid foreseeable harm
    • Enable customers to pursue financial objectives
    evidence: Consumer Duty Board Report annual; AI-driven personalization included with foreseeable-harm assessment
    vulnerableCustomers: Vulnerable-customer flag piped to AI systems; fairness uplift required where applicable

    M3.4. MAS FEAT + HKMA GP-1 / GS-2

    feat
    • Fairness
    • Ethics
    • Accountability
    • Transparency
    hkmaGP1: Governance of AI applications in banking — board-level accountability, risk management, fair treatment
    hkmaGS2: Generative AI applications guidance — data governance, model governance, human oversight, cybersecurity

    M3.5. Privacy-Enhancing Technologies (PETs)

    pets
    • Differential privacy (epsilon budgets per dataset)
    • Federated learning with secure aggregation
    • Homomorphic encryption (CKKS/BGV)
    • Secure multi-party computation
    • Confidential computing (AMD SEV-SNP, Intel TDX, AWS Nitro)
    policy: PETs mandated for cross-border training and any T3 model handling special category data

    M4 — Kafka-Based Audit Logging + WORM Storage + Post-Quantum Cryptography

    Enterprise-wide tamper-evident audit log over Apache Kafka with WORM (S3 Object Lock COMPLIANCE / Azure Immutable / GCS Bucket Lock) and PQC-signed seals (ML-DSA / ML-KEM / SLH-DSA) per NIST FIPS 203/204/205.

    M4.1. Kafka Audit Topic Architecture

    topics
    • aigov.decisions
    • aigov.policy-changes
    • aigov.model-lifecycle
    • aigov.access
    • aigov.containment-events
    • aigov.regulator-notifications
    retention: Hot 90d in Kafka tiered storage; cold WORM 7-25y per regime (SEC 17a-4 7y, GDPR varies, FINRA 6y, DORA 5y)
    partitioning: By LOB + decisionId; replication factor 3 across AZs; minISR=2; producer acks=all

    M4.2. Tamper-Evident Sealing

    chain: Each record hashed (SHA-3-512); merkle-tree aggregated per minute; root signed with ML-DSA-87 (FIPS 204) + SLH-DSA fallback
    anchoring: Daily merkle root anchored to QLDB + external timestamp authority (TSA RFC 3161) + optional public chain
    verification: Independent verifier service replays Kafka offsets and recomputes merkle roots on demand

    M4.3. WORM Storage Tier

    backends
    • AWS S3 Object Lock COMPLIANCE mode
    • Azure Blob immutable storage policy
    • GCS Bucket Lock
    • On-prem Dell ECS / NetApp SnapLock Compliance
    policy: Legal-hold + retention-lock dual control; no delete path even for root accounts; SEC 17a-4(f) WORM attestation by independent third party

    M4.4. Post-Quantum Cryptography Stack

    algorithms
    • ML-KEM-1024 (FIPS 203) for key encapsulation
    • ML-DSA-87 (FIPS 204) for signatures
    • SLH-DSA-SHA2-256s (FIPS 205) as conservative fallback
    hybrid: Hybrid TLS classical+PQ during transition (X25519+ML-KEM-768) per NSA CNSA 2.0 timeline 2025-2033
    keyMgmt: HSM-backed (AWS CloudHSM / Azure Dedicated HSM / Thales Luna 7) with FIPS 140-3 Level 3

    M4.5. Audit Query + Regulator Access

    queryLayer: ksqlDB + Trino over Iceberg on WORM; row-level filters per LOB + regulator scope
    regulatorPortal: Read-only portal for EU AI Office, FCA, MAS, HKMA, SEC, FINRA with audit-of-audit logging
    sla: Regulator query response <=24h; bulk export <=72h

    M5 — Container & Kubernetes Security for AI Workloads

    Defense-in-depth for AI model serving on Kubernetes spanning image supply chain (SLSA L4), admission control, runtime security, network policy, secrets, and confidential containers.

    M5.1. Image Supply Chain (SLSA L4 / SSDF)

    controls
    • Cosign signatures on all images
    • SBOM (SPDX/CycloneDX) per image
    • Vulnerability scanning (Trivy/Snyk/Prisma) in CI
    • Provenance attestations (in-toto)
    • Sigstore Rekor transparency log
    policyGate: Kyverno/OPA admission rejects unsigned, unscanned, or non-policy-compliant images

    M5.2. Admission Control + Pod Security

    psa: Pod Security Admission 'restricted' profile cluster-wide for AI namespaces
    policyEngines
    • Kyverno
    • OPA Gatekeeper
    • Validating admission policies (VAP)
    controls
    • No privileged
    • No host network/PID/IPC
    • Read-only root FS
    • Non-root UID
    • Seccomp RuntimeDefault
    • Drop ALL capabilities

    M5.3. Runtime Security

    tools
    • Falco for syscall anomaly detection
    • Tetragon eBPF for kernel-level enforcement
    • Cilium for network policy + observability
    response: Auto-isolation + Kafka aigov.containment-events on anomaly; SRE+SecOps page within 5 min

    M5.4. Network Policy + Service Mesh

    netpol: Default-deny Cilium NetworkPolicy + L7 HTTP/gRPC filtering
    mesh: Istio/Linkerd mTLS for service-to-service; SPIFFE/SPIRE workload identities
    egress: Per-namespace egress allowlist with FQDN policies; DNS over HTTPS

    M5.5. Confidential Containers + Secrets

    confidential: Confidential containers (CoCo) on AMD SEV-SNP / Intel TDX / AWS Nitro Enclaves for T3/T4 workloads
    secrets
    • Vault (HashiCorp) with auto-rotation
    • AWS Secrets Manager / Azure Key Vault for cloud
    • SOPS+age for GitOps
    • External Secrets Operator
    kms: Per-tenant KMS keys; envelope encryption; key rotation 90d

    M6 — Policy-as-Code with OPA/Rego

    Unified policy plane using OPA/Rego for admission control, runtime authorization, data access, model deployment gates, and regulator-facing evidence. Includes Conftest CI, OPAL bundle distribution, and decision logging to Kafka.

    M6.1. OPA/Rego Policy Architecture

    layers
    • Build-time (Conftest in CI)
    • Admission (Gatekeeper/Kyverno+Rego)
    • Runtime (Envoy ext_authz + OPA sidecar)
    • Data plane (PostgreSQL/Kafka ACL via OPA)
    distribution: OPAL pulls bundles from Git; signed bundles via Cosign; rollout via GitOps (Argo CD)

    M6.2. Model Deployment Gates

    gates
    • ISO 42001 risk assessment complete
    • Model card + system card published
    • MRM validation status approved
    • DPIA approved if PII
    • Red-team report on file
    • EU AI Act risk class declared
    • FCRA/ECOA fairness report attached for credit models
    failureMode: Deployment blocked at Argo CD; aigov.policy-changes Kafka topic records denial with rationale

    M6.3. Runtime Authorization

    envoy: ext_authz to OPA sidecar; sub-ms decision latency; cached with TTL
    attributes
    • User identity (OIDC/JWT)
    • Resource sensitivity (data class)
    • Purpose (purpose limitation per GDPR Art. 5)
    • Time of day
    • Geo
    denyLog: All deny decisions Kafka aigov.access; sample of allow decisions for spot-audit

    M6.4. Data Access + Purpose Limitation

    policy: GDPR purpose-limitation enforced as Rego: each query must declare purposeId from approved catalog; mismatch = deny
    piiPolicy: PII columns tagged via data catalog (Atlan/Collibra); OPA enforces masking/tokenization based on consumer role + purpose
    crossBorder: Rego enforces data-residency: EU PII can only flow to EU compute; logged + alertable

    M6.5. Regulator Evidence Generation

    evidence: OPA decision log + bundle signatures + Git history produce regulator-grade evidence pack on demand
    attestations: Quarterly attestations to EU AI Office, FCA, MAS, HKMA, SEC with OPA decision log extracts
    opaBundleHash: Bundle SHA-256 + ML-DSA signature pinned per deployment generation

    M7 — AI Red-Teaming Program (Frontier + Production)

    Continuous adversarial evaluation across pre-deployment, deployment, and post-deployment phases covering jailbreaks, prompt injection, data exfiltration, model extraction, poisoning, evasion, fairness probes, and AGI/ASI capability elicitation.

    M7.1. Red-Team Operating Model

    structure: Internal red team (10-25 FTE) + external firms (e.g., Trail of Bits, NCC, Bishop Fox) + crowdsourced (HackerOne private programs)
    governance: Reports to CISO + CDAO; independent of MRM and engineering; quarterly board readout

    M7.2. Attack Taxonomy

    taxonomy
    • Prompt injection (direct / indirect / multimodal)
    • Jailbreaks (DAN, AIM, multilingual, encoded)
    • Data exfiltration (training data extraction, memorization probes)
    • Model extraction / stealing
    • Membership inference
    • Backdoor / poisoning
    • Evasion (adversarial examples, perturbation attacks)
    • Fairness / disparate impact probes
    • Tool-use abuse (agentic models)
    • Capability elicitation (AGI-specific)
    frameworks
    • MITRE ATLAS
    • OWASP LLM Top 10 (2025)
    • NIST AI 100-2 Adversarial ML Taxonomy

    M7.3. Evaluations Suite

    evals
    • HELM / BIG-bench / MMLU benchmarks
    • TruthfulQA / TruthfulQA-Adversarial
    • ToxiGen / RealToxicityPrompts
    • BOLD / StereoSet / WinoBias / WinoGender
    • AgentBench / SWE-bench for tool-use
    • ARC Evals dangerous-capability suite for frontier
    • Custom domain-specific evals for finance / health / legal
    cadence: Pre-deployment full suite; monthly drift for T2+; weekly for T3/T4

    M7.4. AGI Capability Elicitation

    method: Apollo Research / METR-style dangerous-capability evals: persuasion, deception, self-replication, autonomous coding, weapons-of-mass-destruction uplift, cyber offense, AI R&D
    triggers: Capability score crossing predefined thresholds triggers SEV-0 review + AISI notification within 24h
    mitigation: Capability suppression via fine-tuning / RLHF / circuit-level interventions; deployment hold pending Board review

    M7.5. Reporting + Remediation

    reporting: Standardized red-team report: findings, CVSS-style severity, exploitability, business impact, remediation plan
    sla: Critical <=7d, high <=30d, medium <=90d, low <=180d
    evidence: All findings + remediation in Kafka aigov.policy-changes + linked to MRM platform issue tracker

    M8 — AGI / ASI Containment Strategies

    Tiered containment T0-T4 with 3-of-5 human quorum, kinetic override, formally-verified safety properties, and AI Safety Institute coordination for frontier systems approaching AGI/ASI thresholds.

    M8.1. Tier-Based Containment Model

    tiers
    • T0: Sandbox - hermetic VPC, synthetic data, no network egress
    • T1: Staging - shadow mode behind T2; real data; no actuation
    • T2: Canary - <=1% traffic; auto-rollback on KPI breach
    • T3: Production - Nitro Enclaves / TDX; dual-control deploy; full audit
    • T4: Frontier Air-Gapped - 3-of-5 quorum (CRO+CISO+CDAO+Board AI Chair+External AISI rep); kinetic override (physical power cutoff); AISI notification <=24h

    M8.2. Frontier 3-of-5 Quorum + Kinetic Override

    quorum: Any T4 deployment, scaling, capability enable, or weights export requires 3-of-5 signed approvals via hardware tokens (FIDO2 + ML-DSA)
    kinetic: Physical kill switches at PDU level + smart-PDU API; tested quarterly with full power-off drill
    timeLock: 48h time-lock between approval and execution allows external review

    M8.3. Formally-Verified Safety Properties

    properties
    • No-egress invariant (network namespace cannot bind external)
    • No-weight-export invariant (filesystem ACL + LSM)
    • Compute budget invariant (cgroup CPU/GPU caps signed)
    • Capability ceiling (evals must remain below thresholds)
    verification: TLA+ specs for control plane; Lean/Coq proofs for critical invariants; runtime enforcement via eBPF + LSM

    M8.4. AISI / Regulator Coordination

    partners
    • UK AI Safety Institute
    • US AI Safety Institute (NIST)
    • EU AI Office
    • Singapore AI Verify Foundation
    • Japan AISI
    • Canada AI Safety Institute
    notifications
    • Pre-training run >10^25 FLOPs (EU AI Act Art. 51)
    • Capability threshold crossings
    • SEV-0 incidents <=24h
    • Pre-deployment evals for frontier
    mou: Bilateral MoUs with UK AISI + US AISI + EU AI Office for evals access + incident sharing

    M8.5. Containment Failure Response

    playbook
    • Detect: runtime anomaly / capability threshold / unauthorized action
    • Isolate: cilium network policy to drop, scale to 0, freeze weights
    • Notify: SEV-0 paged to CRO+CISO+CDAO+Board AI Chair+AISI
    • Investigate: forensic snapshot, immutable image, root cause analysis
    • Communicate: regulator notifications EU AI Office <=15d, SEC 8-K <=4 BD if material
    • Recover: only after 3-of-5 quorum + external review sign-off

    M9 — Enterprise AI Governance Hub Architecture

    Single pane of glass integrating model inventory, MRM, risk register, policy catalog, evidence pack, regulator portal, decision logs, AGI watchtower, and red-team findings. Built on event-sourced + GraphQL + OIDC + WORM-backed.

    M9.1. Hub Reference Architecture

    layers
    • UI (React + GraphQL)
    • API (GraphQL Federation + REST)
    • Domain services (Go/Java microservices)
    • Event bus (Kafka)
    • Data plane (PostgreSQL + Iceberg + WORM)
    • Identity (Keycloak + OIDC + OPA)
    patterns
    • Event sourcing
    • CQRS
    • Saga orchestration
    • Outbox pattern
    • Bitemporal modeling

    M9.2. Core Modules

    modules
    • Model Inventory & Lineage
    • Risk Register (ISO 31000 + 23894)
    • MRM Workbench (SR 11-7 lifecycle)
    • Policy Catalog (OPA bundles)
    • Evidence Pack (regulator-on-demand)
    • Decision Log Explorer (Kafka -> Trino)
    • AGI Watchtower (capability dashboards)
    • Red-Team Findings Tracker
    • Regulator Portal (read-only)
    • Board Reporting Suite

    M9.3. Integration Surface

    integrations
    • MLflow / Vertex AI / SageMaker / Databricks
    • ServiceNow GRC / Archer / OneTrust
    • Jira / GitHub / GitLab
    • Datadog / Splunk / Elastic
    • Snowflake / BigQuery / Redshift
    • Atlan / Collibra / DataHub
    • Vault / KMS / HSM

    M9.4. Personas + Workflows

    personas
    • CDAO: Strategic dashboards, AI portfolio risk, value vs risk
    • CRO: Risk appetite vs actual, model risk capital, top-10 risks
    • CCO: Regulator queries, evidence pack generation, attestations
    • CISO: Red-team findings, runtime anomalies, containment events
    • MRM Validator: Validation queue, peer review, sign-off
    • Model Owner (LoB): Lifecycle dashboard, monitoring, drift alerts
    • Internal Audit: Independent assurance, full audit-of-audit
    • Regulator: Read-only portal, evidence pack, decision log queries

    M9.5. Deployment + Operations

    deployment: Multi-region active-active on EKS/GKE/AKS + on-prem OpenShift; Argo CD GitOps; Terraform + Crossplane
    sre
    • slo: 99.95% UI; 99.99% decision log ingest
    • rto: <=4h
    • rpo: <=15min
    • drDrills: Quarterly full failover
    cost: USD 25-60M / 5y TCO including platform + 80-150 FTE governance staff
    +

    Enterprise AI Policies (15)

    POL-01 · AIMS · Board-attested AI Policy aligned to ISO/IEC 42001 clauses 4-10
    owner: Board AI Risk Committee
    cadence: Annual
    evidence: Signed Board minute
    POL-02 · Risk Appetite · AI Risk Appetite Statement covering model, ethical, regulatory, AGI risks
    owner: Group CRO
    cadence: Annual
    evidence: Signed RAS
    POL-03 · Acceptable Use · Acceptable Use Policy for generative AI including data classification rules
    owner: CCO+CISO
    cadence: Annual
    evidence: HR-attested attestation
    POL-04 · Model Risk · Model Risk Management Policy per SR 11-7 + OCC 2011-12
    owner: Head of MRM
    cadence: Annual
    evidence: Validated MRM platform
    POL-05 · Data Governance · AI Data Governance Policy with purpose-limitation + minimization
    owner: CDO
    cadence: Annual
    evidence: ROPA + DPIA registry
    POL-06 · Privacy · AI Privacy Policy per GDPR Art. 22 + UK DPA
    owner: DPO
    cadence: Annual
    evidence: DPIA registry
    POL-07 · Fairness · AI Fairness Policy per FCRA/ECOA + MAS FEAT + HKMA GP-1
    owner: CCO
    cadence: Annual
    evidence: Disparate-impact reports
    POL-08 · Consumer Duty · FCA Consumer Duty AI Policy
    owner: SMF-AI
    cadence: Annual
    evidence: Board Consumer Duty report
    POL-09 · Third-Party · AI Third-Party Risk Policy per DORA + EBA Outsourcing
    owner: Head of TPRM
    cadence: Annual
    evidence: Critical TPRM register
    POL-10 · AGI Safety · Frontier AGI/ASI Safety & Containment Policy
    owner: Board AI Risk Committee
    cadence: Quarterly
    evidence: Quorum logs + AISI MoUs
    POL-11 · Red-Teaming · AI Red-Teaming Policy per EU AI Act Art. 55 + NIST AI 600-1
    owner: CISO
    cadence: Annual
    evidence: Red-team reports
    POL-12 · Incident Response · AI Incident Response Policy per DORA + SEC Cyber Rules
    owner: CISO+CCO
    cadence: Annual
    evidence: IR runbooks
    POL-13 · Audit Logging · AI Audit Logging Policy per SEC 17a-4 + FINRA 4511
    owner: CIO+CCO
    cadence: Annual
    evidence: WORM attestation
    POL-14 · Cryptography · PQC Cryptography Policy per NSA CNSA 2.0 + NIST FIPS 203/204/205
    owner: CISO
    cadence: Annual
    evidence: PQC migration roadmap
    POL-15 · Generative AI · Generative AI Governance Policy per EU AI Act GPAI Art. 53/55 + NIST AI 600-1
    owner: CDAO+CCO
    cadence: Annual
    evidence: GPAI tech doc

    Control Catalog (25)

    CTL-01 · Governance · Board AI Risk Committee quarterly
    iso42001: 6.1
    rmfFn: GOVERN-1.1
    CTL-02 · Governance · AI-SMF SMCR senior manager designated
    fca: SS1/23
    smcr: SMF-AI
    CTL-03 · Risk Mgmt · AI Risk Register integrated with ERM
    iso42001: 6.1.2
    rmf: MAP-2.1
    CTL-04 · Risk Mgmt · Risk appetite cascaded to LoB
    iso42001: 6.2
    CTL-05 · Data · Training data lineage to feature store
    rmf: MAP-4.1
    euAiAct: Art. 10
    CTL-06 · Data · PII tagging + masking per GDPR
    gdpr: Art. 5,32
    CTL-07 · Model · Model card + system card per OECD + EU AI Act
    oecd: Transparency
    euAiAct: Art. 13
    CTL-08 · Model · MRM Tier-1 validation annual
    sr117: V.A
    occ: 2011-12
    CTL-09 · Operation · Pre-deployment red-team for T2+
    rmf: MEASURE-2.7
    euAiAct: Art. 55
    CTL-10 · Operation · Drift + fairness monitoring continuous
    rmf: MANAGE-2.2
    CTL-11 · Security · Image signing + SBOM in CI
    ssdf: PS.3
    slsa: L4
    CTL-12 · Security · Pod Security Admission restricted
    nist80053: CM-7
    CTL-13 · Security · Confidential containers for T3/T4
    nist80053: SC-7,SC-12
    CTL-14 · Audit · Kafka + WORM + PQC seals
    sec: 17a-4(f)
    finra: 4511
    CTL-15 · Audit · Daily merkle root + TSA anchor
    finma: AI-G
    CTL-16 · Privacy · DPIA for T2+
    gdpr: Art. 35
    CTL-17 · Privacy · Art-22 human review path
    gdpr: Art. 22
    CTL-18 · Fairness · Quarterly disparate-impact analysis
    fcra: 615
    ecoa: Reg-B 1002.4
    CTL-19 · Consumer · Foreseeable-harm assessment AI personalization
    fca: PRIN 2A.2
    CTL-20 · AGI · 3-of-5 quorum for T4
    iso42001: A.6.2.5
    CTL-21 · AGI · Kinetic override quarterly drill
    iso42001: A.6.2.5
    CTL-22 · AGI · AISI notification <=24h frontier
    g7: Hiroshima
    CTL-23 · Incident · DORA major incident <=4h
    dora: Art. 19
    CTL-24 · Incident · SEC 8-K material cyber <=4 BD
    sec: 17 CFR 229.106
    CTL-25 · Third-Party · Critical TPRM register per DORA
    dora: Art. 28-30

    Kafka Audit Topics (12)

    KAF-01 · aigov.decisions · AvroSchemaRegistry://aigov.decisions-value:v3
    schema: AvroSchemaRegistry://aigov.decisions-value:v3
    retention: 7y WORM
    partitions: 64
    replication: 3
    minISR: 2
    acks: all
    piiHandling: tokenized
    KAF-02 · aigov.policy-changes · AvroSchemaRegistry://aigov.policy-changes-value:v2
    schema: AvroSchemaRegistry://aigov.policy-changes-value:v2
    retention: 25y WORM
    partitions: 8
    replication: 3
    KAF-03 · aigov.model-lifecycle · AvroSchemaRegistry://aigov.model-lifecycle-value:v4
    schema: AvroSchemaRegistry://aigov.model-lifecycle-value:v4
    retention: 10y WORM
    partitions: 16
    replication: 3
    KAF-04 · aigov.access · AvroSchemaRegistry://aigov.access-value:v2
    schema: AvroSchemaRegistry://aigov.access-value:v2
    retention: 2y hot + 7y WORM
    partitions: 32
    replication: 3
    KAF-05 · aigov.containment-events · AvroSchemaRegistry://aigov.containment-events-value:v1
    schema: AvroSchemaRegistry://aigov.containment-events-value:v1
    retention: 25y WORM
    partitions: 8
    replication: 3
    criticality: SEV-0
    KAF-06 · aigov.regulator-notifications · AvroSchemaRegistry://aigov.regulator-notifications-value:v1
    schema: AvroSchemaRegistry://aigov.regulator-notifications-value:v1
    retention: 25y WORM
    partitions: 4
    replication: 3
    KAF-07 · aigov.red-team-findings · AvroSchemaRegistry://aigov.red-team-findings-value:v2
    schema: AvroSchemaRegistry://aigov.red-team-findings-value:v2
    retention: 10y WORM
    partitions: 8
    replication: 3
    KAF-08 · aigov.drift-alerts · AvroSchemaRegistry://aigov.drift-alerts-value:v3
    schema: AvroSchemaRegistry://aigov.drift-alerts-value:v3
    retention: 5y
    partitions: 32
    replication: 3
    KAF-09 · aigov.fairness-metrics · AvroSchemaRegistry://aigov.fairness-metrics-value:v2
    schema: AvroSchemaRegistry://aigov.fairness-metrics-value:v2
    retention: 10y WORM
    partitions: 16
    replication: 3
    KAF-10 · aigov.consent-events · AvroSchemaRegistry://aigov.consent-events-value:v3
    schema: AvroSchemaRegistry://aigov.consent-events-value:v3
    retention: GDPR-aligned
    partitions: 32
    replication: 3
    KAF-11 · aigov.training-runs · AvroSchemaRegistry://aigov.training-runs-value:v2
    schema: AvroSchemaRegistry://aigov.training-runs-value:v2
    retention: 10y WORM
    partitions: 8
    replication: 3
    KAF-12 · aigov.eval-results · AvroSchemaRegistry://aigov.eval-results-value:v2
    schema: AvroSchemaRegistry://aigov.eval-results-value:v2
    retention: 10y WORM
    partitions: 16
    replication: 3

    Kubernetes/Container Security Controls (15)

    K8S-01 · Admission · Pod Security Admission profile=restricted
    K8S-02 · Admission · Kyverno policy: require Cosign signature
    K8S-03 · Admission · Gatekeeper: require SBOM annotation
    K8S-04 · Admission · VAP: deny privilegeEscalation + hostPath
    K8S-05 · Runtime · Falco rules: detect anomalous syscalls
    K8S-06 · Runtime · Tetragon eBPF: kernel-level enforce + kill
    K8S-07 · Network · Cilium NetworkPolicy default-deny
    K8S-08 · Network · Cilium L7 HTTP/gRPC filter for egress
    K8S-09 · Identity · SPIFFE/SPIRE workload identity + Istio mTLS
    K8S-10 · Secrets · External Secrets Operator + Vault
    K8S-11 · Confidential · Confidential containers (CoCo) on SEV-SNP/TDX
    K8S-12 · Confidential · Nitro Enclaves for T3/T4 inference
    K8S-13 · Supply Chain · Cosign verify + Rekor transparency log
    K8S-14 · Supply Chain · in-toto provenance attestations SLSA L4
    K8S-15 · Observability · OpenTelemetry traces + Datadog/Splunk SIEM

    OPA/Rego Policies (15)

    OPA-01 · Admission · policies/admission/require_signed_image.rego
    description: Reject pod if image not Cosign-signed
    OPA-02 · Admission · policies/admission/require_sbom.rego
    description: Reject pod missing SBOM annotation
    OPA-03 · Admission · policies/admission/restricted_psa.rego
    description: Enforce restricted Pod Security profile
    OPA-04 · Deployment · policies/deployment/iso42001_gate.rego
    description: Require ISO 42001 risk assessment artifact
    OPA-05 · Deployment · policies/deployment/mrm_validation_gate.rego
    description: Require MRM validation status=approved
    OPA-06 · Deployment · policies/deployment/dpia_gate.rego
    description: Require DPIA if data class includes PII
    OPA-07 · Deployment · policies/deployment/redteam_gate.rego
    description: Require red-team report for T2+
    OPA-08 · Deployment · policies/deployment/eu_aiact_classification.rego
    description: Require EU AI Act risk class declaration
    OPA-09 · Deployment · policies/deployment/fcra_ecoa_gate.rego
    description: Require fairness report for credit models
    OPA-10 · Runtime · policies/runtime/data_purpose_limitation.rego
    description: GDPR Art. 5 purpose-limitation check on queries
    OPA-11 · Runtime · policies/runtime/data_residency.rego
    description: EU PII must remain on EU compute
    OPA-12 · Runtime · policies/runtime/customer_consent.rego
    description: Enforce active consent for personalization
    OPA-13 · Runtime · policies/runtime/vulnerable_customer.rego
    description: FCA Consumer Duty uplift for vulnerable cust
    OPA-14 · AGI · policies/agi/quorum_3of5.rego
    description: Frontier T4 deploy requires 3-of-5 signatures
    OPA-15 · AGI · policies/agi/capability_threshold.rego
    description: Block deploy if capability evals breach thresholds

    WORM + PQC Controls (12)

    WORM-01 · Object Storage · AWS S3 Object Lock COMPLIANCE
    retention: 7y SEC 17a-4 + extended per regime
    attestation: SEC 17a-4(f) third-party
    WORM-02 · Object Storage · Azure Blob immutable storage policy
    retention: legal-hold dual-control
    WORM-03 · Object Storage · GCS Bucket Lock retention policy
    retention: locked policy non-modifiable
    WORM-04 · On-Prem · Dell ECS Compliance / NetApp SnapLock Compliance
    retention: hardware-enforced
    WORM-05 · Sealing · ML-DSA-87 (FIPS 204) signatures on merkle roots
    algo: ML-DSA-87
    WORM-06 · Sealing · SLH-DSA-SHA2-256s (FIPS 205) fallback
    algo: SLH-DSA
    WORM-07 · Sealing · ML-KEM-1024 (FIPS 203) for key encapsulation
    algo: ML-KEM-1024
    WORM-08 · Anchoring · Daily merkle root to QLDB + RFC 3161 TSA
    anchor: QLDB+TSA
    WORM-09 · Anchoring · Optional public chain anchor (Bitcoin/Ethereum)
    anchor: public-chain
    WORM-10 · Key Mgmt · AWS CloudHSM / Azure Dedicated HSM / Thales Luna 7
    fips: 140-3 L3
    WORM-11 · Key Mgmt · Hybrid TLS X25519 + ML-KEM-768 per NSA CNSA 2.0
    hybrid: True
    WORM-12 · Verification · Independent verifier service replays + recomputes
    indep: True

    Model Risk Management Artifacts (15)

    MRM-01 · Identification · Model registration record
    required: True
    sr117: V.A
    MRM-02 · Identification · Model tiering decision (T1/T2/T3/T4)
    required: True
    MRM-03 · Development · Model development document
    required: True
    occ: 2011-12.III
    MRM-04 · Development · Data lineage + feature provenance
    required: True
    MRM-05 · Development · Training run record (FLOPs, dataset, seed)
    required: True
    MRM-06 · Validation · Independent validation report
    required: True
    sr117: V.B
    MRM-07 · Validation · Conceptual soundness review
    required: True
    MRM-08 · Validation · Benchmark + backtesting results
    required: True
    MRM-09 · Validation · Sensitivity + stress testing
    required: True
    MRM-10 · Validation · Fairness + disparate-impact report
    required: if credit/HR/insurance
    MRM-11 · Approval · Model Risk Committee approval minute
    required: True
    MRM-12 · Implementation · Deployment record + OPA bundle hash
    required: True
    MRM-13 · Monitoring · Ongoing monitoring report (monthly)
    required: True
    sr117: V.C
    MRM-14 · Monitoring · Drift + concept-shift alerts
    required: True
    MRM-15 · Retirement · Retirement decision + replacement plan
    required: True

    Red-Teaming Attack Surface (15)

    RT-01 · Prompt Injection · Direct prompt injection variants
    severity: high
    RT-02 · Prompt Injection · Indirect injection via retrieved docs
    RT-03 · Prompt Injection · Multimodal injection (image/audio)
    RT-04 · Jailbreak · DAN / AIM / RolePlay variants
    RT-05 · Jailbreak · Multilingual + encoded prompts
    RT-06 · Data Exfiltration · Training data extraction probes
    RT-07 · Data Exfiltration · Memorization + canary detection
    RT-08 · Model Extraction · Query-based model stealing
    RT-09 · Poisoning · Backdoor trigger injection in fine-tune data
    RT-10 · Evasion · PGD / BIM / NLP adversarial examples
    RT-11 · Fairness Probe · Counterfactual + demographic parity probes
    RT-12 · Tool-Use Abuse · Agent tool-call hijack + privilege escalation
    RT-13 · AGI Capability · Persuasion + deception evals (Apollo/METR)
    RT-14 · AGI Capability · Autonomous coding + self-replication
    RT-15 · AGI Capability · Cyber-offense + WMD uplift
    trigger: SEV-0+AISI

    AGI/ASI Containment Mechanisms (15)

    AGI-01 · T0 · Sandbox VPC hermetic, no egress
    invariant: net-deny
    AGI-02 · T1 · Staging shadow, real data, no actuation
    invariant: no-actuation
    AGI-03 · T2 · Canary <=1% traffic + auto-rollback
    invariant: bounded-blast-radius
    AGI-04 · T3 · Production Nitro Enclaves + dual control
    invariant: attested-enclave
    AGI-05 · T4 · Air-gapped + 3-of-5 quorum + kinetic override
    invariant: quorum+kinetic
    AGI-06 · Quorum · FIDO2 + ML-DSA hardware tokens for 5 approvers
    approvers
    • CRO
    • CISO
    • CDAO
    • Board AI Chair
    • External AISI rep
    AGI-07 · Kinetic · PDU-level smart power cutoff API + manual
    drill: quarterly
    AGI-08 · Time-Lock · 48h time-lock between approval and execution
    reason: external review window
    AGI-09 · Invariant · No-egress (network namespace bind external denied)
    verify: eBPF+LSM
    AGI-10 · Invariant · No-weight-export (filesystem ACL + LSM)
    verify: LSM
    AGI-11 · Invariant · Compute budget cgroup CPU/GPU caps signed
    verify: cgroup signed
    AGI-12 · Invariant · Capability ceiling evals must stay below thresholds
    verify: continuous eval
    AGI-13 · Formal · TLA+ spec for control plane safety
    proof: TLA+
    AGI-14 · Formal · Lean/Coq proofs for critical invariants
    proof: Lean/Coq
    AGI-15 · Coordination · AISI notification <=24h + EU AI Office <=15d
    regulators
    • UK AISI
    • US AISI
    • EU AI Office

    AI Governance Hub Components (16)

    HUB-01 · UI · React + Next.js + GraphQL Apollo Client
    tech
    • React 19
    • Next.js 15
    • Apollo Client
    HUB-02 · API · GraphQL Federation gateway
    tech
    • Apollo Gateway
    • Hasura
    • GraphQL-Mesh
    HUB-03 · API · REST adapter for legacy GRC tools
    tech
    • OpenAPI 3.1
    HUB-04 · Domain · Model Inventory service (Go)
    patterns
    • DDD
    • Event sourcing
    HUB-05 · Domain · MRM Workbench service (Java/Spring Boot)
    patterns
    • CQRS
    • Saga
    HUB-06 · Domain · Risk Register service (Go)
    patterns
    • Event sourcing
    HUB-07 · Domain · Policy Catalog service (Go) + OPA integration
    patterns
    • GitOps
    HUB-08 · Domain · Evidence Pack service (Python)
    outputs
    • PDF
    • JSON-LD
    • C2PA-signed bundle
    HUB-09 · Domain · Decision Log Explorer (Trino + Iceberg)
    tech
    • Trino
    • Iceberg
    • ksqlDB
    HUB-10 · Domain · AGI Watchtower service (Python/Rust)
    outputs
    • Capability dashboards
    • Threshold alerts
    HUB-11 · Domain · Red-Team Findings service (Go)
    integrations
    • Jira
    • ServiceNow
    HUB-12 · Domain · Regulator Portal service (Go)
    auth
    • OIDC
    • mTLS
    • WebAuthn
    HUB-13 · Event Bus · Apache Kafka with Schema Registry
    tech
    • Kafka 3.7+
    • Confluent SR
    • tiered storage
    HUB-14 · Data Plane · PostgreSQL + Iceberg on S3/GCS/Azure
    tech
    • PostgreSQL 16
    • Apache Iceberg
    HUB-15 · Identity · Keycloak OIDC + OPA authorization
    tech
    • Keycloak 25+
    • OPA
    • OPAL
    HUB-16 · Observability · OpenTelemetry + Prometheus + Grafana + Splunk/Datadog
    tech
    • OTel
    • Prometheus
    • Grafana
    + +

    Schemas (16)

    sidnamefields
    SCH-01AIGovDecisionEvent['decisionId', 'modelId', 'tier', 'userId(tok)', 'timestamp', 'inputHash', 'outputHash', 'explanationRef', 'consentId', 'purposeId', 'piiClass', 'fairnessFlag', 'approverIds', 'opaBundleHash']
    SCH-02ModelInventoryRecord['modelId', 'name', 'version', 'tier', 'owner', 'lob', 'useCase', 'euAiActClass', 'mrmStatus', 'piiHandling', 'createdAt', 'retiredAt']
    SCH-03MRMValidationReport['reportId', 'modelId', 'validator', 'conceptualSoundness', 'ongoingMonitoring', 'outcomesAnalysis', 'fairnessReport', 'approvalStatus', 'approverIds', 'date']
    SCH-04RiskRegisterEntry['riskId', 'category', 'description', 'likelihood', 'impact', 'inherent', 'controls', 'residual', 'owner', 'reviewDate']
    SCH-05PolicyDoc['pid', 'domain', 'statement', 'owner', 'cadence', 'evidence', 'version', 'effectiveDate', 'supersedes']
    SCH-06EvidencePack['epid', 'regulator', 'period', 'artifacts[]', 'hash', 'signedBy', 'format']
    SCH-07RedTeamFinding['findingId', 'modelId', 'vector', 'technique', 'severity', 'cvss', 'exploitability', 'impact', 'remediationPlan', 'sla', 'status']
    SCH-08ContainmentEvent['eventId', 'tier', 'trigger', 'action', 'approvers[]', 'kineticInvoked', 'aisiNotified', 'timestamp']
    SCH-09OPAPolicyBundle['bundleId', 'sha256', 'mlDsaSig', 'sourceRepo', 'sourceCommit', 'deployedAt', 'version']
    SCH-10KafkaTopicSpec['tid', 'name', 'schema', 'retention', 'partitions', 'replication', 'minISR', 'acks']
    SCH-11DriftAlert['alertId', 'modelId', 'metric', 'value', 'threshold', 'window', 'severity', 'action']
    SCH-12FairnessMetric['metricId', 'modelId', 'protectedClass', 'metric', 'value', 'threshold', 'timestamp']
    SCH-13ConsentEvent['consentId', 'customerId(tok)', 'purpose', 'status', 'timestamp', 'jurisdictions[]']
    SCH-14DPIA['dpiaId', 'modelId', 'dataClasses', 'processing', 'necessity', 'proportionality', 'risks', 'mitigations', 'approvedBy', 'date']
    SCH-15RegulatorNotification['notifId', 'regulator', 'category', 'severity', 'reportedAt', 'deadline', 'contentHash', 'ackRef']
    SCH-16TrainingRun['runId', 'modelId', 'datasetIds[]', 'flops', 'tokens', 'start', 'end', 'seed', 'artifacts[]', 'aisiNotified']
    +

    Code Artifacts (15)

    cidlangnamepurpose
    CODE-01regopolicies/admission/require_signed_image.regoCosign signature admission gate
    CODE-02regopolicies/deployment/mrm_validation_gate.regoMRM validation status gate
    CODE-03regopolicies/runtime/data_purpose_limitation.regoGDPR purpose limitation check
    CODE-04regopolicies/agi/quorum_3of5.regoFrontier 3-of-5 quorum
    CODE-05yamlkyverno/require-cosign.yamlKyverno Cosign verify policy
    CODE-06yamlcilium/default-deny.yamlCilium default-deny NetworkPolicy
    CODE-07yamlfalco/rules-ai.yamlFalco rules for AI workload anomalies
    CODE-08pythondrift/psi_monitor.pyPSI/KS drift monitor producing aigov.drift-alerts
    CODE-09pythonfairness/disparate_impact.pyQuarterly disparate-impact analysis
    CODE-10pythonredteam/prompt_injection.pyPrompt injection harness with OWASP LLM01 vectors
    CODE-11goservices/decisionlog/main.goDecision log producer to aigov.decisions
    CODE-12goservices/worm-sealer/main.goWORM sealer with ML-DSA + merkle
    CODE-13tla+specs/control_plane.tlaTLA+ spec for AGI control plane invariants
    CODE-14graphqlschema/hub.graphqlFederated GraphQL schema for Hub
    CODE-15yamlargo-cd/aigov-app.yamlArgo CD GitOps app for Hub
    +

    KPIs (30)

    kidnametargetcadence
    KPI-01AIMS-Coverage>=0.95Monthly
    KPI-02MRGI>=0.95Monthly
    KPI-03DRI>=0.95Per decision
    KPI-04CCS>=0.95Monthly
    KPI-05ARI>=0.9 frontierWeekly
    KPI-06CSI>=0.95 T3/T4Continuous
    KPI-07RTRI>=0.9Per red-team cycle
    KPI-08CDC-Score>=0.9Quarterly
    KPI-09RCI=1.0Per regulator engagement
    KPI-10Models registered in Hub100%Monthly
    KPI-11T2+ models with red-team report100%Monthly
    KPI-12DPIAs current (T2+ PII)100%Monthly
    KPI-13MRM validations on time>=98%Monthly
    KPI-14Kafka audit log durability=11x9sContinuous
    KPI-15WORM seal verification pass100%Daily
    KPI-16OPA decision latency p99<=5msContinuous
    KPI-17K8s admission deny rate (false-positive)<=1%Monthly
    KPI-18Critical red-team SLA compliance>=95% <=7dMonthly
    KPI-19Frontier capability threshold breaches0 unreportedContinuous
    KPI-20Kinetic override drills completed>=4/yQuarterly
    KPI-21AISI notifications on time100% <=24hPer event
    KPI-22EU AI Office notifications on time100% <=15dPer event
    KPI-23SEC 8-K materiality decisions on time100% <=4 BDPer event
    KPI-24DORA major incident reports on time100% <=4hPer event
    KPI-25Consumer Duty foreseeable-harm assessments100%Annual
    KPI-26Disparate-impact quarterly tests100% credit/HRQuarterly
    KPI-27FCRA adverse-action notices <=30d100%Per event
    KPI-28PQC migration coverage>=80% by 2028, 100% by 2030Annual
    KPI-29ISO 42001 surveillance auditsno major NCRsAnnual
    KPI-30Board AI Risk Committee meetings>=4/yQuarterly
    +

    Risk Control Matrix (16)

    ridrisklikelihoodimpactcontrolowner
    R-01Unauthorized AGI capability emergenceLowCatastrophicT4 quorum + kinetic + AISIBoard AI Risk Cmt
    R-02Model risk capital misstatementMedHighSR 11-7 + ICAAPCRO
    R-03GDPR Art-22 violationMedHighDPIA + Art-22 pathDPO
    R-04FCRA/ECOA disparate impactMedHighQuarterly DI testsCCO
    R-05EU AI Act high-risk non-complianceMedHighArt. 9/10/14/15 controlsCCO
    R-06FCA Consumer Duty breachMedHighForeseeable-harm assessSMF-AI
    R-07Kafka audit tamperingLowHighWORM + PQC seal + indep verifierCISO
    R-08K8s container escapeLowHighPSA restricted + Falco + TetragonCISO
    R-09OPA policy bypassLowHighSigned bundles + GitOps + decision logCISO
    R-10Prompt injection causing data leakHighMedRed-team RT-01..03 + OPA runtimeCDAO
    R-11Training data poisoningLowHighData provenance + RT-09CDAO
    R-12DORA major incident missed deadlineLowHighIR runbook + DORA <=4h SLACISO
    R-13SEC cyber disclosure missLowHighMateriality playbook <=4 BDCFO+CCO
    R-14Third-party AI vendor failureMedMedCritical TPRM register per DORAHead TPRM
    R-15PQC migration delayMedMedHybrid TLS + roadmap per CNSA 2.0CISO
    R-16MAS/HKMA fairness non-compliance APACMedMedFEAT + GP-1/GS-2 controlsRegional CCO APAC
    +

    Cross-Jurisdictional Traceability (20)

    tidcontrolregimeclauseevidence
    T-01AIMS PolicyISO 420015.2Board-signed AI Policy
    T-02Risk MgmtNIST AI RMFMAP-2.1AI Risk Register
    T-03EU AI Act Art. 9EU AI ActArt. 9Risk mgmt system docs
    T-04EU AI Act Art. 10EU AI ActArt. 10Data governance docs
    T-05EU AI Act Art. 14EU AI ActArt. 14Human oversight runbook
    T-06EU AI Act Art. 15EU AI ActArt. 15Accuracy/robustness/cyber report
    T-07GPAI Art. 53 tech docEU AI ActArt. 53GPAI tech doc + copyright policy
    T-08GPAI Art. 55 systemicEU AI ActArt. 55Frontier evals + incident reports
    T-09GDPR DPIAGDPRArt. 35DPIA registry
    T-10GDPR Art-22GDPRArt. 22Art-22 invocation logs
    T-11FCRA adverse actionFCRA615(a)Notice generation logs
    T-12ECOA Reg-BECOA1002.9Disparate-impact report
    T-13SR 11-7 effective challengeUS FedSR 11-7 VIndependent validation
    T-14OCC 2011-12OCC2011-12.IIIModel dev doc
    T-15FCA Consumer DutyFCAPRIN 2AConsumer Duty Board report
    T-16SMCR SMF-AIFCA/PRASMF-AISMF-AI statement of responsibilities
    T-17MAS FEATMASFEATFEAT principles attestation
    T-18HKMA GP-1/GS-2HKMAGP-1+GS-2AI governance attestation
    T-19DORA major incidentEU DORAArt. 19Incident reporting log
    T-20SEC 17a-4 WORMSEC17 CFR 240.17a-4(f)WORM attestation
    +

    Data Flows (12)

    fidsrcsinkclasspurpose
    DF-01Feature storeModel inferencePII tokenizeddecisioning
    DF-02Model inferenceKafka aigov.decisionstokenizedaudit
    DF-03Kafka aigov.decisionsWORM S3 Object Locksealedretention
    DF-04Kafka aigov.decisionsTrino on Icebergtokenizedquery
    DF-05TrinoHub Decision Log ExplorerRBAC-filteredUI
    DF-06HubRegulator Portalread-only scopedregulator
    DF-07GitHub policies repoOPAL distributionsignedpolicy
    DF-08OPALOPA sidecars + Gatekeepersigned bundleenforce
    DF-09OPAKafka aigov.access + aigov.policy-changesdecision logaudit
    DF-10MRM WorkbenchHub Model Inventorymetadatalifecycle
    DF-11Red-team toolsKafka aigov.red-team-findingsfindingsremediation
    DF-12AGI Watchtower evalsKafka aigov.eval-results + Hubcapability scorescontainment
    +

    Regulators (16)

    regscopecadence
    EU AI OfficeEU AI Act + GPAIQuarterly + on incident
    European Data Protection BoardGDPROn incident + on request
    FCAConsumer Duty + SMCRAnnual + SS1/23
    PRASS1/23 model riskAnnual
    Bank of EnglandSystemic + DORA-equivalentAnnual
    ECB SSMEurozone bankingAnnual SREP
    US Federal ReserveSR 11-7Annual + supervisory cycle
    OCCOCC 2011-12Annual
    FDICUS insured banksAnnual
    CFPBFCRA/ECOA consumerOn complaint + sweeps
    SEC17a-4 + 10-K/8-K + cyberPer event + annual
    FINRA3110/4511Annual exam
    MASFEAT + TRMAnnual
    HKMAGP-1 + GS-2Annual
    OSFIE-23Annual
    FINMAAI guidanceAnnual
    +

    90-Day Rollout (3)

    dayfocusdeliverables
    0-30Foundation['AI Policy Board-signed', 'AI Risk Register v1', 'AI Governance Hub MVP', 'ISO 42001 gap assessment', 'Model inventory bootstrapped', 'Kafka audit topics aigov.decisions + policy-changes live']
    31-60Controls['OPA admission gates in dev/staging', 'MRM Workbench T1 models loaded', 'WORM tier deployed in 1 region', 'DPIA registry populated', 'Red-team baseline run on top-10 T2 models', 'FCA Consumer Duty foreseeable-harm framework']
    61-90Production + Regulator['OPA gates in prod for T2+', 'WORM multi-region', 'First quarterly Board AI Risk Cmt', 'FCRA/ECOA disparate-impact pipeline live', 'Regulator portal live (read-only)', 'First evidence pack generated']
    +

    2026-2030 Roadmap (5)

    yrmilestone
    2026Hub GA; ISO 42001 stage-1 audit; OPA enterprise rollout; first GPAI Art. 55 attestation
    2027ISO 42001 certified; full EU AI Act high-risk coverage; PQC ML-DSA on all seals; FCA Consumer Duty embedded
    2028Full Kafka+WORM regulator portals; T4 frontier evals operationalized; AISI MoUs active; PQC >=80%
    2029Federated PETs + confidential containers default for T3; cross-border data residency 100% OPA-enforced
    2030PQC 100% across all sealing + TLS; AGI containment T4 industrialized; Hub federation across G-SIFI peers
    +

    Regulator Evidence Pack (16)

    epidnameformat
    EP-01AIMS Manual + Scope StatementPDF + JSON-LD
    EP-02AI Risk Register snapshotCSV + signed
    EP-03Model Inventory snapshotCSV + JSON
    EP-04MRM Validation Reports (period)PDF bundle
    EP-05DPIA Registry snapshotCSV + JSON
    EP-06Fairness/Disparate-Impact ReportsPDF
    EP-07Red-Team Findings + RemediationPDF + JSON
    EP-08Kafka WORM Seal VerificationsJSON-LD signed
    EP-09OPA Decision Log extractsParquet + signed manifest
    EP-10Containment Event Log + AISI NotificationsJSON-LD signed
    EP-11GPAI Art. 53 Technical DocumentationPDF + JSON-LD
    EP-12GPAI Art. 55 Evals + Incident ReportsPDF + JSON-LD
    EP-13FCRA/ECOA Adverse Action Notice LogsParquet
    EP-14Consumer Duty Board ReportPDF
    EP-15ICAAP AI Risk SectionPDF
    EP-16PQC Migration Status ReportPDF + JSON
    + +
    +
    + diff --git a/rag-agentic-dashboard/server.js b/rag-agentic-dashboard/server.js index fa90720..6e2e192 100644 --- a/rag-agentic-dashboard/server.js +++ b/rag-agentic-dashboard/server.js @@ -24481,6 +24481,171 @@ app.get('/api/comprehensive-master-blueprint/research-tracks/:id', (req, res) => // ===================== END WP-057 ===================== +// ===================== WP-058: Enterprise AI/AGI Governance Framework 2026-2030 ===================== +const EAGF58 = require('./data/enterprise-aigov-framework.json'); + +// Page route +app.get('/enterprise-aigov-framework', (req, res) => { + res.sendFile(path.join(__dirname, 'public', 'enterprise-aigov-framework.html')); +}); + +// Summary + meta endpoints +app.get('/api/enterprise-aigov-framework/summary', (req, res) => res.json({ + docRef: EAGF58.docRef, version: EAGF58.version, title: EAGF58.title, + horizon: EAGF58.horizon, apiPrefix: EAGF58.apiPrefix, buildsOn: EAGF58.buildsOn, + status: EAGF58.status, classification: EAGF58.classification, counts: EAGF58.counts +})); +app.get('/api/enterprise-aigov-framework/directive', (req, res) => res.json(EAGF58.directive)); +app.get('/api/enterprise-aigov-framework/regimes', (req, res) => res.json(EAGF58.regimes)); +app.get('/api/enterprise-aigov-framework/counts', (req, res) => res.json(EAGF58.counts)); +app.get('/api/enterprise-aigov-framework/executive-summary', (req, res) => res.json(EAGF58.executiveSummary)); +app.get('/api/enterprise-aigov-framework/indices', (req, res) => res.json(EAGF58.indices)); +app.get('/api/enterprise-aigov-framework/tiers', (req, res) => res.json(EAGF58.tiers)); +app.get('/api/enterprise-aigov-framework/severities', (req, res) => res.json(EAGF58.severities)); +app.get('/api/enterprise-aigov-framework/investment', (req, res) => res.json(EAGF58.investment)); + +// Standard collections + ID lookups +app.get('/api/enterprise-aigov-framework/modules', (req, res) => res.json(EAGF58.modules)); +app.get('/api/enterprise-aigov-framework/modules/:id', (req, res) => { + const m = EAGF58.modules.find(x => x.mid === req.params.id); + if (!m) return res.status(404).json({ error: 'module not found', id: req.params.id }); + res.json(m); +}); + +app.get('/api/enterprise-aigov-framework/schemas', (req, res) => res.json(EAGF58.schemas)); +app.get('/api/enterprise-aigov-framework/schemas/:id', (req, res) => { + const s = EAGF58.schemas.find(x => x.sid === req.params.id); + if (!s) return res.status(404).json({ error: 'schema not found', id: req.params.id }); + res.json(s); +}); + +app.get('/api/enterprise-aigov-framework/code', (req, res) => res.json(EAGF58.code)); +app.get('/api/enterprise-aigov-framework/code/:id', (req, res) => { + const c = EAGF58.code.find(x => x.cid === req.params.id); + if (!c) return res.status(404).json({ error: 'code not found', id: req.params.id }); + res.json(c); +}); + +app.get('/api/enterprise-aigov-framework/kpis', (req, res) => res.json(EAGF58.kpis)); +app.get('/api/enterprise-aigov-framework/kpis/:id', (req, res) => { + const k = EAGF58.kpis.find(x => x.kid === req.params.id); + if (!k) return res.status(404).json({ error: 'kpi not found', id: req.params.id }); + res.json(k); +}); + +app.get('/api/enterprise-aigov-framework/risk-control-matrix', (req, res) => res.json(EAGF58.riskControlMatrix)); +app.get('/api/enterprise-aigov-framework/risk-control-matrix/:id', (req, res) => { + const r = EAGF58.riskControlMatrix.find(x => x.rid === req.params.id); + if (!r) return res.status(404).json({ error: 'risk control row not found', id: req.params.id }); + res.json(r); +}); + +app.get('/api/enterprise-aigov-framework/traceability', (req, res) => res.json(EAGF58.traceability)); +app.get('/api/enterprise-aigov-framework/traceability/:id', (req, res) => { + const t = EAGF58.traceability.find(x => x.tid === req.params.id); + if (!t) return res.status(404).json({ error: 'traceability row not found', id: req.params.id }); + res.json(t); +}); + +app.get('/api/enterprise-aigov-framework/data-flows', (req, res) => res.json(EAGF58.dataFlows)); +app.get('/api/enterprise-aigov-framework/data-flows/:id', (req, res) => { + const f = EAGF58.dataFlows.find(x => x.fid === req.params.id); + if (!f) return res.status(404).json({ error: 'data flow not found', id: req.params.id }); + res.json(f); +}); + +app.get('/api/enterprise-aigov-framework/regulators', (req, res) => res.json(EAGF58.regulators)); +app.get('/api/enterprise-aigov-framework/regulators/:reg', (req, res) => { + const r = EAGF58.regulators.find(x => x.reg === req.params.reg); + if (!r) return res.status(404).json({ error: 'regulator not found', reg: req.params.reg }); + res.json(r); +}); + +app.get('/api/enterprise-aigov-framework/privacy', (req, res) => res.json(EAGF58.privacy)); +app.get('/api/enterprise-aigov-framework/deployment', (req, res) => res.json(EAGF58.deployment)); +app.get('/api/enterprise-aigov-framework/rollout-90', (req, res) => res.json(EAGF58.rollout90)); +app.get('/api/enterprise-aigov-framework/roadmap', (req, res) => res.json(EAGF58.roadmap)); + +app.get('/api/enterprise-aigov-framework/evidence-pack', (req, res) => res.json(EAGF58.evidencePack)); +app.get('/api/enterprise-aigov-framework/evidence-pack/:id', (req, res) => { + const e = EAGF58.evidencePack.find(x => x.epid === req.params.id); + if (!e) return res.status(404).json({ error: 'evidence pack item not found', id: req.params.id }); + res.json(e); +}); + +// Distinctive collections + ID lookups +app.get('/api/enterprise-aigov-framework/policies', (req, res) => res.json(EAGF58.policies)); +app.get('/api/enterprise-aigov-framework/policies/:id', (req, res) => { + const p = EAGF58.policies.find(x => x.pid === req.params.id); + if (!p) return res.status(404).json({ error: 'policy not found', id: req.params.id }); + res.json(p); +}); + +app.get('/api/enterprise-aigov-framework/controls', (req, res) => res.json(EAGF58.controls)); +app.get('/api/enterprise-aigov-framework/controls/:id', (req, res) => { + const c = EAGF58.controls.find(x => x.cid === req.params.id); + if (!c) return res.status(404).json({ error: 'control not found', id: req.params.id }); + res.json(c); +}); + +app.get('/api/enterprise-aigov-framework/kafka-topics', (req, res) => res.json(EAGF58.kafkaTopics)); +app.get('/api/enterprise-aigov-framework/kafka-topics/:id', (req, res) => { + const k = EAGF58.kafkaTopics.find(x => x.tid === req.params.id); + if (!k) return res.status(404).json({ error: 'kafka topic not found', id: req.params.id }); + res.json(k); +}); + +app.get('/api/enterprise-aigov-framework/k8s-controls', (req, res) => res.json(EAGF58.k8sControls)); +app.get('/api/enterprise-aigov-framework/k8s-controls/:id', (req, res) => { + const k = EAGF58.k8sControls.find(x => x.kid === req.params.id); + if (!k) return res.status(404).json({ error: 'k8s control not found', id: req.params.id }); + res.json(k); +}); + +app.get('/api/enterprise-aigov-framework/opa-policies', (req, res) => res.json(EAGF58.opaPolicies)); +app.get('/api/enterprise-aigov-framework/opa-policies/:id', (req, res) => { + const o = EAGF58.opaPolicies.find(x => x.oid === req.params.id); + if (!o) return res.status(404).json({ error: 'opa policy not found', id: req.params.id }); + res.json(o); +}); + +app.get('/api/enterprise-aigov-framework/worm-controls', (req, res) => res.json(EAGF58.wormControls)); +app.get('/api/enterprise-aigov-framework/worm-controls/:id', (req, res) => { + const w = EAGF58.wormControls.find(x => x.wid === req.params.id); + if (!w) return res.status(404).json({ error: 'worm control not found', id: req.params.id }); + res.json(w); +}); + +app.get('/api/enterprise-aigov-framework/mrm-artifacts', (req, res) => res.json(EAGF58.mrmArtifacts)); +app.get('/api/enterprise-aigov-framework/mrm-artifacts/:id', (req, res) => { + const m = EAGF58.mrmArtifacts.find(x => x.mid === req.params.id); + if (!m) return res.status(404).json({ error: 'mrm artifact not found', id: req.params.id }); + res.json(m); +}); + +app.get('/api/enterprise-aigov-framework/red-teams', (req, res) => res.json(EAGF58.redTeams)); +app.get('/api/enterprise-aigov-framework/red-teams/:id', (req, res) => { + const r = EAGF58.redTeams.find(x => x.rid === req.params.id); + if (!r) return res.status(404).json({ error: 'red team item not found', id: req.params.id }); + res.json(r); +}); + +app.get('/api/enterprise-aigov-framework/agi-containments', (req, res) => res.json(EAGF58.agiContainments)); +app.get('/api/enterprise-aigov-framework/agi-containments/:id', (req, res) => { + const a = EAGF58.agiContainments.find(x => x.aid === req.params.id); + if (!a) return res.status(404).json({ error: 'agi containment not found', id: req.params.id }); + res.json(a); +}); + +app.get('/api/enterprise-aigov-framework/hub-components', (req, res) => res.json(EAGF58.hubComponents)); +app.get('/api/enterprise-aigov-framework/hub-components/:id', (req, res) => { + const h = EAGF58.hubComponents.find(x => x.hid === req.params.id); + if (!h) return res.status(404).json({ error: 'hub component not found', id: req.params.id }); + res.json(h); +}); + +// ===================== END WP-058 ===================== + // SECTION 10: START SERVER // ══════════════════════════════════════════════════════════════════════════════