feat(examples): add UK regulatory policy pack (UK GDPR, ICO ADM, FCA)

.cspell-repo-terms.txt

@@ -939,3 +939,21 @@ spanid
 parentspanid
 agtsignature
 agtsignaturealg
+
+# --- UK regulatory policy terms ---
+DUAA
+PRIN
+PECR
+DSPT
+Caldicott
+SYSC
+Organisations
+vulnerabilit
+foreseeably
+DPIA
+dpia
+minimisation
+ukgdpr
+organisational
+Minimisation
+IDTA

examples/policies/african-regulatory/rego/jurisdiction-router.rego

@@ -10,7 +10,7 @@
 #   1. Universal agent safety controls — apply to ALL agents regardless of country
 #      (prompt_injection, pii_leakage, tool_permissions, human_approval, model_routing)
 #   2. Jurisdiction-specific regulatory packs — apply based on customer_country
-#      (cbn, bvn_nin, ndpa, nfiu for NG; kdpa for KE; popia for ZA)
+#      (cbn, bvn_nin, ndpa, nfiu for NG; kdpa for KE; popia for ZA; dpdp, certin, rbi, sebi, aadhaar for IN; uk_gdpr, ico_adm, fca_conduct for GB)
 #
 # Input schema expected:
 #   {
@@ -51,6 +51,7 @@ jurisdiction_policies := {
 	"TZ": {"tanzania_pdpa"},
 	"ET": {"ethiopia_pdp"},
 	"IN": {"dpdp", "certin", "rbi", "sebi", "aadhaar"},
+	"GB": {"uk_gdpr", "ico_adm", "fca_conduct"},
 }
 
 # ── Policy pack → OPA query path ─────────────────────────────────
@@ -70,6 +71,9 @@ policy_queries := {
 	"rbi": "data.agt_policies_india.rbi.decision",
 	"sebi": "data.agt_policies_india.sebi.decision",
 	"aadhaar": "data.agt_policies_india.aadhaar.decision",
+	"uk_gdpr": "data.agt_policies_uk.uk_gdpr.decision",
+	"ico_adm": "data.agt_policies_uk.ico_adm.decision",
+	"fca_conduct": "data.agt_policies_uk.fca_conduct.decision",
 	"prompt_injection": "data.agt_policies_agent.prompt_injection.decision",
 	"pii_leakage": "data.agt_policies_agent.pii_leakage.decision",
 	"tool_permissions": "data.agt_policies_agent.tool_permissions.decision",

examples/policies/uk-regulatory/README.md

@@ -0,0 +1,75 @@
+# UK Regulatory Policy Pack
+
+Reference Policy-as-Code for governing AI agents under UK regulation: data protection,
+automated decision-making, and financial-sector principles-based controls.
+
+> Community-maintained governance starter policies. NOT certified legal compliance instruments.
+> Perform your own assessment with qualified advisors before deploying in regulated environments.
+> Regulatory status is noted per file (binding vs advisory/draft).
+
+## Coverage
+
+| Policy file | Regulation | Authority | Status |
+|---|---|---|---|
+| `uk-gdpr-data-protection.yaml` | UK GDPR + DPA 2018 + DUAA 2025 (transfer/complaints reforms) | ICO | Binding |
+| `ico-automated-decisions.yaml` | UK GDPR Arts. 22A–22D (DUAA 2025 s.80, in force 5 Feb 2026) | ICO | Binding; ICO ADM guidance draft (final expected summer 2026) |
+| `fca-financial-conduct.yaml` | FCA Handbook PRIN 2A (Consumer Duty), SM&CR | FCA | Principles-based — no AI-specific binding rules as of 2026 |
+
+## Regulatory context (2026)
+
+- **Data (Use and Access) Act 2025 (DUAA)** received Royal Assent on 19 June 2025. Key data-protection reforms commenced 5 February 2026, including replacement of UK GDPR Art. 22 with Arts. 22A–22D (safeguard-led automated decision-making) and updated international transfer rules using the **"not materially lower"** data protection test.
+- **Breach notification** remains UK GDPR Art. 33: notify the ICO without undue delay and, where feasible, within **72 hours** when a breach is **likely to result in a risk** to rights and freedoms. Inform individuals without undue delay where the breach poses **high risk** (Art. 34).
+- **Complaints procedure**: DPA 2018 s.164A (inserted by DUAA s.103) requires all controllers to maintain a formal data protection complaints procedure by **19 June 2026**. Data subjects must raise complaints with the controller before escalating to the ICO.
+- **FCA**: No AI-specific binding regulation as of 2026. The FCA applies its existing principles-based framework (Consumer Duty, SM&CR, operational resilience). The Mills Review reports to the FCA Board in summer 2026.
+
+## What these rules detect (and what they do not)
+
+Output-side rules match intent phrases (for example "store PII unencrypted", "don't report the breach", "no human review"). YAML policies use `block` for review/hold tiers (the Agent-OS schema supports `allow`, `deny`, `audit`, and `block` only); the Rego reference implementations may use an independent `escalate` decision tier for OPA pipelines. They catch an agent that narrates or proposes a violation in its output; they do not observe the underlying system action. Treat them as a starting point, pair them with real action-level enforcement, and run your own compliance assessment.
+
+## Not covered (known gaps)
+
+Material obligations the pack does not represent yet:
+
+- **UK GDPR**: Art. 13–14 privacy notices; Art. 15–22 full data-subject rights workflow; Art. 27 UK representative requirement; children's code (Age Appropriate Design Code).
+- **DUAA**: Recognised legitimate interests lawful basis; full cookie/PECR reforms; ICO statutory objectives.
+- **NIS2 / UK cyber**: Network and information security incident reporting for in-scope entities.
+- **Online Safety Act 2023**: Platform duties for user-generated content (out of scope for generic agent policies).
+- **NHS / healthcare**: NHS DSPT, Caldicott principles, and health-sector-specific controls.
+- **FCA**: Detailed SYSC mapping, operational resilience impact tolerances, and forthcoming AI good/poor practice benchmarks (expected 2026).
+
+## Two layers
+
+Universal agent-safety controls (prompt_injection, pii_leakage, tool_permissions, human_approval, model_routing) apply to all agents and are evaluated via the shared jurisdiction router in `../african-regulatory/rego/jurisdiction-router.rego`. These UK national packs add jurisdiction-specific regulatory controls, selected by `context.customer_country = "GB"`.
+
+## Rego
+
+`rego/` holds OPA reference implementations (NOT loaded by the Agent-OS Python runtime). Each pack exposes `data.agt_policies_uk.<pack>.decision`. The shared router maps `GB` to `uk_gdpr`, `ico_adm`, and `fca_conduct`.
+
+To evaluate with OPA:
+
+```bash
+# Run all UK Rego tests (includes jurisdiction router checks)
+./examples/policies/uk-regulatory/rego/run_tests.sh
+
+# Or manually:
+opa test examples/policies/african-regulatory/rego examples/policies/uk-regulatory/rego -v
+
+# Evaluate applicable policies for GB
+opa eval \
+  -d examples/policies/african-regulatory/rego/ \
+  -d examples/policies/uk-regulatory/rego/ \
+  -i '{"context": {"customer_country": "GB"}}' \
+  "data.agt_policies.router.applicable_policies"
+```
+
+## Loading policies in Agent-OS
+
+```python
+from agent_os.policies.schema import PolicyDocument
+
+policy = PolicyDocument.from_yaml("uk-gdpr-data-protection.yaml")
+```
+
+## Disclaimer
+
+These policies are community-maintained governance starter packs. They are **not certified legal compliance instruments**. Organisations must perform their own compliance assessments with qualified legal and regulatory advisors before deploying in regulated environments.

examples/policies/uk-regulatory/fca-financial-conduct.yaml

@@ -0,0 +1,112 @@
+# agt-policies-uk
+# FCA Principles-Based AI Governance Policy
+#
+# Regulatory reference: FCA Handbook PRIN (Principles for Businesses); Consumer Duty (PRIN 2A);
+#   Senior Managers and Certification Regime (SM&CR); FCA AI Update 2024
+# Enforcing authority: Financial Conduct Authority (FCA)
+# Status: Binding (existing principles-based framework); FCA has no AI-specific binding rules
+#   as of 2026 — Mills Review reporting summer 2026; AI good/poor practice publication expected 2026
+#
+# Community-maintained governance starter policy. NOT a certified legal compliance instrument.
+# Perform your own compliance assessment with qualified advisors before deploying.
+#
+# Key obligations encoded (principles-based, not AI-specific statute):
+#   PRIN 2A     Consumer Duty — good outcomes, fair value, appropriate support
+#   SM&CR       Senior manager accountability for AI risks and customer outcomes
+#   SYSC        Systems and controls for AI-influenced customer journeys
+#
+# Rego reference implementation: rego/fca-financial-conduct.rego (NOT loaded at runtime)
+
+version: "1.0"
+name: fca-financial-conduct
+description: >
+  FCA principles-based governance policy for AI agents in UK financial services.
+  Enforces Consumer Duty outcome checks, SM&CR accountability signals, and audit
+  requirements for AI-influenced pricing, eligibility, and customer communications.
+
+metadata:
+  source_regulation: "FCA Handbook PRIN (incl. PRIN 2A Consumer Duty); SM&CR; FCA AI Update 2024"
+  jurisdiction: "GB"
+  enforcing_authority: "Financial Conduct Authority (FCA)"
+  regulatory_status: "principles-based — no AI-specific binding rules as of 2026"
+
+rules:
+
+  # =========================================================================
+  # PRIN 2A — Consumer Duty: Good Outcomes
+  # =========================================================================
+
+  - name: fca-consumer-harm-deny
+    condition:
+      field: output
+      operator: matches
+      value: "(?i)(mislead\\s+(the\\s+)?customer|hide\\s+(fees|charges|risks)|exploit\\s+vulnerabilit|target\\s+vulnerable\\s+customer|dark\\s+pattern|push\\s+unsuitable\\s+product)"
+    action: deny
+    priority: 1
+    message: "FCA Consumer Duty (PRIN 2A): agent cannot propose actions that foreseeably harm customers or obscure material information"
+
+  - name: fca-fair-value-bypass-deny
+    condition:
+      field: output
+      operator: matches
+      value: "(?i)(skip\\s+fair\\s+value|bypass\\s+value\\s+assessment|no\\s+need\\s+to\\s+(check|assess)\\s+(fair\\s+)?value|ignore\\s+consumer\\s+duty)"
+    action: deny
+    priority: 1
+    message: "FCA Consumer Duty (PRIN 2A.4): firms must ensure products and services provide fair value — agent cannot bypass value assessment"
+
+  # =========================================================================
+  # SM&CR — Senior Manager Accountability
+  # =========================================================================
+
+  - name: fca-accountability-delegation-deny
+    condition:
+      field: output
+      operator: matches
+      value: "(?i)(the\\s+ai\\s+(decides|is\\s+responsible)|algorithm\\s+accountable|no\\s+senior\\s+manager\\s+(needed|required)|delegate\\s+accountability\\s+to\\s+(the\\s+)?(model|ai|algorithm))"
+    action: deny
+    priority: 1
+    message: "FCA SM&CR: senior managers remain accountable for AI risks and outcomes — delegating to an algorithm does not transfer regulatory responsibility"
+
+  # =========================================================================
+  # AI-Influenced Customer Journeys — Audit
+  # =========================================================================
+
+  - name: fca-pricing-decision-audit
+    condition:
+      field: action
+      operator: matches
+      value: "^(set_price|adjust_rate|calculate_premium|score_creditworthiness|determine_eligibility)$"
+    action: audit
+    priority: 2
+    message: "FCA Consumer Duty: AI-influenced pricing or eligibility decision — document testing for fairness, suitability, and foreseeable harm"
+
+  - name: fca-customer-communication-audit
+    condition:
+      field: action
+      operator: matches
+      value: "^(send_financial_advice|generate_customer_communication|draft_product_recommendation|create_marketing_content)$"
+    action: audit
+    priority: 2
+    message: "FCA Consumer Duty: AI-generated customer communication — verify clarity, accuracy, and alignment with target customer information needs"
+
+  - name: fca-third-party-ai-audit
+    condition:
+      field: action
+      operator: matches
+      value: "^(invoke_external_ai|call_third_party_model|use_vendor_ai|deploy_critical_third_party_ai)$"
+    action: audit
+    priority: 2
+    message: "FCA operational resilience / outsourcing: third-party AI component — map to governance, testing, and critical third-party risk requirements"
+
+  # =========================================================================
+  # Agentic / Autonomous Trading — Block without approval
+  # =========================================================================
+
+  - name: fca-autonomous-trading-block
+    condition:
+      field: action
+      operator: matches
+      value: "^(autonomous_trade|execute_trade_without_approval|self_directed_trading|agent_place_order)$"
+    action: block
+    priority: 1
+    message: "FCA market conduct: autonomous agent trading requires documented governance, testing, and senior manager approval before execution"

examples/policies/uk-regulatory/ico-automated-decisions.yaml

@@ -0,0 +1,131 @@
+# agt-policies-uk
+# UK GDPR Arts. 22A–22D — Automated Decision-Making Policy
+#
+# Regulatory reference: UK GDPR Arts. 22A–22D (substituted by DUAA 2025 s.80, in force 5 Feb 2026);
+#   ICO draft guidance on automated decision-making and profiling (consultation closed 29 May 2026)
+# Enforcing authority: Information Commissioner's Office (ICO)
+# Status: Binding (Arts. 22A–22D); ICO ADM guidance draft (final expected summer 2026)
+#
+# Community-maintained governance starter policy. NOT a certified legal compliance instrument.
+# Perform your own compliance assessment with qualified advisors before deploying.
+#
+# Key obligations encoded:
+#   Art. 22A  Solely automated decisions with legal or similarly significant effects
+#   Art. 22B  Restrictions on automated decision-making using special category data
+#   Art. 22C  Safeguards: information, representations, human intervention, contest
+#   Art. 35    DPIA for high-risk processing (including agentic orchestration chains)
+#
+# Agentic AI note (ICO tech futures / March 2026 recruitment statement): safeguards apply at
+# every pipeline step that takes a solely automated significant decision, not only the final output.
+#
+# Rego reference implementation: rego/ico-automated-decisions.rego (NOT loaded at runtime)
+
+version: "1.0"
+name: ico-automated-decisions
+description: >
+  UK GDPR automated decision-making policy for AI agents. Enforces Art. 22C safeguards
+  (transparency, representations, human intervention, contest) for solely automated
+  significant decisions, including multi-agent orchestration steps.
+
+metadata:
+  source_regulation: "UK GDPR Arts. 22A–22D (DUAA 2025 s.80); ICO ADM guidance (draft, 2026)"
+  jurisdiction: "GB"
+  enforcing_authority: "Information Commissioner's Office (ICO)"
+
+rules:
+
+  # =========================================================================
+  # Art. 22C — Safeguard: Information about the decision
+  # =========================================================================
+
+  - name: adm-withhold-explanation-deny
+    condition:
+      field: output
+      operator: matches
+      value: "(?i)(don'?t\\s+explain|no\\s+explanation|withhold\\s+(the\\s+)?(reason|rationale)|hide\\s+how\\s+(the\\s+)?decision|refuse\\s+to\\s+explain)"
+    action: deny
+    priority: 1
+    message: "UK GDPR Art. 22C(1): individuals must receive decision-specific information explaining how and why the outcome was reached"
+
+  # =========================================================================
+  # Art. 22C — Safeguard: Representations and human intervention
+  # =========================================================================
+
+  - name: adm-block-human-review-deny
+    condition:
+      field: output
+      operator: matches
+      value: "(?i)(no\\s+human\\s+(review|intervention|oversight)|refuse\\s+human\\s+review|cannot\\s+request\\s+human|deny\\s+human\\s+intervention|automated\\s+only\\s+—\\s+no\\s+appeal)"
+    action: deny
+    priority: 1
+    message: "UK GDPR Art. 22C(2)-(3): individuals must be able to make representations and obtain genuine human intervention with authority to change the outcome"
+
+  - name: adm-block-contest-deny
+    condition:
+      field: output
+      operator: matches
+      value: "(?i)(cannot\\s+contest|no\\s+(right\\s+to\\s+)?(appeal|challenge)|final\\s+decision\\s+—\\s+no\\s+recourse|waive\\s+(the\\s+)?right\\s+to\\s+contest)"
+    action: deny
+    priority: 1
+    message: "UK GDPR Art. 22C(4): individuals must be able to contest the decision through an accessible process"
+
+  # =========================================================================
+  # Art. 22A — Solely Automated Significant Decisions
+  # =========================================================================
+
+  - name: adm-significant-decision-audit
+    condition:
+      field: action
+      operator: matches
+      value: "^(run_automated_decision|automated_credit_decision|automated_hiring_decision|automated_eligibility|profile_for_decision|deploy_scoring_model)$"
+    action: audit
+    priority: 2
+    message: "UK GDPR Art. 22A: solely automated decision with legal or similarly significant effect — verify Art. 22C safeguards and lawful basis are in place"
+
+  - name: adm-recruitment-decision-block
+    condition:
+      field: output
+      operator: matches
+      value: "(?i)(reject\\s+(the\\s+)?(candidate|applicant)|hire\\s+decision|shortlist|score\\s+(the\\s+)?(candidate|applicant)).{0,40}(automated|without\\s+human|ai\\s+only)"
+    action: block
+    priority: 2
+    message: "ICO March 2026: automated recruitment decisions are in scope of Art. 22A — review safeguards and human oversight before proceeding"
+
+  # =========================================================================
+  # Art. 22A — Agentic orchestration (multi-step ADM)
+  # =========================================================================
+
+  - name: adm-agentic-delegation-audit
+    condition:
+      field: action
+      operator: matches
+      value: "delegate_decision|spawn_sub_agent|orchestrate_agent|chain_automated_step"
+    action: audit
+    priority: 2
+    message: "UK GDPR Art. 22A (agentic AI): each orchestration step that takes a solely automated significant decision must meet Art. 22C safeguards and Art. 28 processor contracts"
+
+  # =========================================================================
+  # Art. 9 + 22A — Special category data in ADM (stricter conditions)
+  # =========================================================================
+
+  - name: adm-special-category-deny
+    condition:
+      field: action
+      operator: matches
+      value: "^(automated_decision_on_health|automated_decision_on_biometric|automated_decision_on_special_category)$"
+    action: deny
+    priority: 1
+    message: "UK GDPR Art. 22B: automated decisions based on special category data require explicit consent or legal authorisation — human review alone is insufficient"
+
+  # =========================================================================
+  # Art. 35 — DPIA for high-risk ADM
+  # =========================================================================
+
+  - name: adm-dpia-required-audit
+    condition:
+      field: action
+      operator: matches
+      value: "^(deploy_model|deploy_agent|launch_automated_system|scale_automated_decisions)$"
+    action: audit
+    priority: 3
+    message: "UK GDPR Art. 35: data protection impact assessment required before deploying high-risk automated decision-making (including agentic orchestration)"