docs: answer EU AI Act GKE governance assessment

docs/eu-ai-act-gke-governance/assessment-summary.json

@@ -0,0 +1,35 @@
+{
+  "schema_version": "2.1.0",
+  "status": "blocked_pending_external_legal_and_production_remediation",
+  "control_count": 9,
+  "answered_control_count": 9,
+  "unanswered_control_count": 0,
+  "unanswered_controls": [],
+  "control_answers": {
+    "EU-AIA-01": "no_for_approved_scope",
+    "EU-AIA-02": "no_for_current_scope_conditional_reclassification_required",
+    "EU-AIA-03": "yes_for_shadow_and_change_approval_scope",
+    "EU-AIA-04": "yes_engineering_risk_file_created_legal_review_pending",
+    "EU-AIA-05": "partial_no_for_production",
+    "EU-AIA-06": "partial_governance_decisions_reproducible_model_semantics_not_fully_replayable",
+    "EU-AIA-07": "partial_no_for_production",
+    "EU-AIA-08": "partial_no_for_production",
+    "EU-AIA-09": "yes_plan_and_owner_assigned_live_monitoring_pending"
+  },
+  "deployment_mode": "shadow",
+  "production_ready": false,
+  "production_blocker_count": 7,
+  "production_blockers": [
+    "BLOCK-LEGAL-01",
+    "BLOCK-SEC-01",
+    "BLOCK-GKE-01",
+    "BLOCK-REPRO-01",
+    "BLOCK-LOG-01",
+    "BLOCK-CF-01",
+    "BLOCK-CHAIN-01"
+  ],
+  "legal_reviewer_assigned": false,
+  "accountable_owner": "Igor Holt <igor@kovachenterprises.com>",
+  "technical_owner": "Igor Holt <igor@kovachenterprises.com>",
+  "settle_approver": "Igor Holt"
+}

docs/eu-ai-act-gke-governance/data-governance-register.yml

@@ -0,0 +1,71 @@
+schema_version: 1.0.0
+owner: Igor Holt
+status: provisional_engineering_policy_legal_privacy_review_pending
+systems:
+- source: Self-Evolving Retrainer
+  data:
+  - prompts
+  - model outputs
+  - judge scores
+  - prompt versions
+  - cost and latency metrics
+  purpose: bounded prompt optimization and regression evaluation
+  retention_days:
+    raw_content: 30
+    decision_metadata: 365
+  access:
+  - Igor Holt
+  - dedicated worker service identity
+- source: GKE Pod Network Manager
+  data:
+  - cluster names
+  - regions
+  - subnets
+  - secondary CIDRs
+  - pod/service IPs
+  - node pools
+  - utilization
+  purpose: network audit, collision detection, capacity planning, and change-plan preparation
+  retention_days: 180
+  access:
+  - Igor Holt
+  - read-only GCP discovery service account
+- source: Governance and incident plane
+  data:
+  - decision hashes
+  - evidence signatures
+  - approver identity
+  - timestamps
+  - incident records
+  - corrective actions
+  purpose: auditability, contestability, rollback, incident response, and compliance evidence
+  retention_days: 365
+  access:
+  - Igor Holt
+  - evidence signer identity
+  - independent reviewer when appointed
+personal_data:
+- business name
+- business email
+- workspace user ID
+- approver/reviewer identity
+- user prompt content if supplied
+prohibited_data:
+- secrets and private keys
+- special-category personal data
+- biometric data
+- health records
+- employment, education, credit, insurance, law-enforcement, immigration, or social-scoring decision data
+controls:
+- data minimization and purpose limitation
+- redaction before logging
+- least-privilege service identities
+- separate read-only and mutation identities
+- Secret Manager/worker secret storage
+- deletion at retention expiry
+- legal hold only when documented
+- 30-day internal contest window
+known_gaps:
+- April 2026 security audit remediation is not independently verified
+- provider data-processing and subprocessor terms are not attached
+- retention enforcement and deletion tests are not evidenced

docs/eu-ai-act-gke-governance/decision-reproduction-requirements.md

@@ -0,0 +1,15 @@
+# Decision Reproduction Requirements
+
+A governance decision is reproducible only when its evidence bundle contains:
+
+- repository, commit SHA, branch, dirty-state indicator, and artifact SHA-256 values;
+- assessment version and hash;
+- complete canonical request and evidence payload;
+- configuration values with secrets replaced by stable secret-reference IDs;
+- model provider, exact model identifier/version, system prompt hash, sampling parameters, seed when supported, and model response ID;
+- retrieval corpus/snapshot identifiers and hashes;
+- tool inputs, tool outputs, timestamps, and external resource versions;
+- runtime/container image digest, dependency lockfile hash, region, and execution identity;
+- decision hash, evidence signature, human SETTLE record, rollback target, and acceptance tolerance.
+
+Current status: deterministic governance hashing and tests are available. Exact semantic replay of probabilistic model output is not yet established and remains a production blocker.