Skip to content

v4.1.0 — Cybernetic Governance Engine

Choose a tag to compare

@lahlfors lahlfors released this 13 Nov 02:05
· 102 commits to main since this release
23b8c39

Release Summary: Corrective verification and compliance-custody release backed by the matched-SHA two-project Gate W certification archive at commit fa7721221d44a53d01def9d810dfbd01c96d4293 (gs://<PROJECT_ID>-evidence-worm-staging/gate-w-certification/fa772122/, archive_manifest_sha256: 70513815ba2cfe7f57a1bf5d5106528cad5f605387307717569bd254e6551be4, compliance/gate_w_certification.json). Ships everything accumulated in [Unreleased] since v4.0.0 (cage-audit/3.0 stream invariants, live repository-ruleset verification, CAGE_INFERENCE_BACKEND, Valkey AOF persistence, model-free kernel/install isolation) together with the full Wave 1–4 and P0–P2 GKE remediation (#675–#730): custom-VPC binding (CP-6), derived data-store /32 egress (DE-1), non-loopback Linkerd mesh identity (LT-1), blocking refusal/deferral evidence commits (BR-1–BR-3), deployment/k8s/ shadow-manifest retirement (KM-1–KM-2), in-cluster advisor proxy routing and GEAP/AGW retirement (AR-1, EA-1–EA-3), human OIDC + WebAuthn HITL quorum binding (HI-1), mesh-authenticated INFRA custody (IC-1–IC-4), server-bound Python STPA enforcement (ST-1), native Kubernetes initContainers OPA UDS sidecar (OP-1–OP-2), per-workload network/mesh segmentation (NS-1–NS-2), persistent 120-day CMEK evidence domain and bound cross-teardown genesis rotation (ED-1), KMS Cloud Audit Log corroboration, in-cluster Lula v0.16.0 CronJob auditing (26/26 active specs satisfied under startup_grace_hours == 0) with a 2 × schedule dead-man's switch, and regional EU_ECB / APAC_MAS scope alignment (5 adopter-operated institutional controls marked not-applicable, gdpr-art22 split between HOLDS runtime enforcement and planned EU-006 contestation intake). Numbered 4.1.0 rather than 5.0.0 because this work makes 4.0.0's existing governance, identity, and evidence-custody claims true on GKE rather than introducing a new major capability generation; see docs/BREAKING_CHANGES.md (IB-1–OP-2) and the v4.0.0 post-release errata below. Ships alongside client-v0.3.0, the first functional release of packages/cage-client.

Added

  • Matched-SHA two-project Gate W certification (fa772122), persistent WORM/KMS custody corroboration, in-cluster Lula v0.16.0 auditor + dead-man's switch, native initContainers OPA sidecar, regional SSP scope alignment, and cage-client v0.3.0 (#675–#730) (gate_w_certification.json, staging_lifecycle.sh, evidence_domain, lula_cron, gateway, policy.py, packages/cage-client, POAM.md).

    • Gate W empirical certification (fa772122, 2026-11-12): Verified at identical main SHA across both a zero-state clean-room GKE project (fresh_project_clean_room, max cold pull+queue 156.247s vs 600s Phase 2 timeout archived in kubelet-image-pull-durations-*.json) and a recovery validation project (fresh_project_with_phase0_kms_residue, exercising tombstoned DELETED WIF pool/provider undelete, --import-state recovery of the Binary Authorization Grafeas note with surviving image occurrences + KMS/WORM set, and read_worm_chain_anchor.py cold-start binding to a prior multi-batch WORM chain): Phase 1 two-stage zero-drift terraform plan, Phase 2 9/9 workloads Ready with restartCount == 0, Phase 3 26/26 active Lula v0.16.0 specs satisfied (0 not-satisfied) under startup_grace_hours == 0, Phase 4 US_FED/EU_ECB/APAC_MAS/residency/Langfuse compliance isolation, Phase 5 live sweeps (live_external, integration, 12/12 test_gke_trade_flow.py preview -> 2-human OIDC+WebAuthn quorum -> POST_HITL execution, safety_rate == 1.0 with blocked_traces == 0 across all 5 controls, GMP cage_* timeSeries, bidirectional KMS Cloud Audit Log corroboration), and Phase 6 intra-run teardown -> rebuild (bound_rotations >= 1, declared_gaps == 0) -> final destroy (0 unexpected survivors). Closes 34 POA&M items with finding-ID-bound regression tests while tracking POAM-2026-201 (lula_audit -> A.5.3 observer-subject feedback spiral) openly.
    • Durable post-W detection & quarterly expiry (#705, #707): CronJob/lula-audit emits structured evaluated_count / satisfied_count / failed_controls telemetry enforced against a >= 26 active-manifest corpus floor; failing runs evaluate to blocking (False) on A.5.3, fire notifier.send_critical_alert(), and pair with a 2 × schedule dead-man's switch (sla_monitor.py + google_monitoring_alert_policy bound to @lahlfors), while scripts/check_gate_w_freshness.py fails scheduled CI once the 90-day certification window expires.
    • Regional SSP & Lula scope alignment (#730): Reclassifies 5 institutional adopter obligations (eu-ai-act-art9, dora-art10, mas-feat, mas-notice-655, mas-trm-s6) to not-applicable (control-origination: adopter-operated) with concrete 3-step adopter integration runbooks, splits gdpr-art22 into an implemented (HOLDS) runtime component (REQUIRE_APPROVAL, distinct-principal quorum, POST_HITL drift revalidation) and a planned controller contestation component (EU-006), keeps ra-5 planned under POAM-010, and refreshes the 209-row OSCAL evidence-claim census.
    • cage-client v0.3.0 (client-v0.3.0, #675, #676): First functional SDK release. Discloses three facts together: (1) client-v0.1.0 and client-v0.2.0 never completed a call against a live gateway (defer_token vs. deferred_id / envelope schema mismatch); (2) out-of-band kid GovernanceEnvelope.signature verification (ES256/EdDSA over RFC 8785 JCS, expires_at freshness, and subject.action_hash binding) is now enforced by default and cannot be disabled in enforcing postures; and (3) preview GovernanceEnvelope dropped the unpopulated subject.record_hash and governance_context.controls_satisfied fields (EP-1).
  • CI compares the repository's live protection with the documented configuration, and POAM-2026-116, POAM-2026-133 and POAM-2026-170 close on it (ci(ci), check_repo_rules.py, repo-rules.yml, test_check_repo_rules.py, branch-protection-rules.md, CONTRIBUTING.md, POAM.md).

    • check_repo_rules.py reads the rules on main and the release-tag ruleset from the GitHub API and fails on any difference from what the document says, in either direction: a rule, check, pin or ruleset missing or disabled, and also one added. A fetch error fails it too. repo-rules.yml runs it on every pull request, every push to main and daily, because what it checks can change without a commit.
    • The tests run seventeen planted drifts against the API's recorded responses and require each to be reported. They also hold the document's required-check table equal to the checker's pins, so the two cannot drift apart. Eight mutants of the checker each fail them.
    • Run live against a branch no ruleset covers, the checker reported all four of main's rules missing and exited 1. Bypass actors are visible only to admins, so a bypass actor added to a ruleset goes unseen.
    • CONTRIBUTING.md had its own protection table and settings list: CODEOWNERS review, protected rc-v* and release/**, seven checks it called required, three of which are not and one of which (Marker Contract Check) does not exist, tag patterns no rule uses, and read-and-write workflow permissions. It now summarises what is enforced and points to .github/branch-protection-rules.md, as JURISDICTION_INVENTORY.md does, so one list of required checks remains.
  • Pure-stdlib cage-audit/3.0 stream-invariant oracle (INV-1–INV-5), live Compose --stream CI gate, and GOVERNANCE_THRESHOLDS_LOADED attestation (POAM-2026-151) (feat(governance), stream_invariants.py, check_trace_conformance.py, thresholds.py, ci.yml, docker-compose.yml).

    • proof/stream_invariants.py (zero src/ imports, pure-stdlib RFC 8785 JCS + SHA-256) verifies cage-audit/3.0 chain integrity, GOVERNANCE_THRESHOLDS_LOADED digest agreement (POAM-2026-151 Option 2 producer, together with thresholds_digest on GET /governance/policy-version), INV-1 actuation seal provenance, INV-2 single-use nonce consumption, INV-3 refusal receipt integrity (proof_hash + payload_hash), INV-4 barrier non-violation, and INV-5 settlement ledger conservation.
    • POST /tools/execute surfaces proof_hash on tool-caught governance refusals, docker-compose.yml provisions an internal redis:7.2-alpine (noeviction) service and defaults local dev signing to CAGE_KMS_PROVIDER=ed25519 (SoftwareEd25519Provider EdDSA JWT seals so CAGE_SEAL_STRICT_MODE=true stays enforced out of the box), QUICKSTART.md adds an explicit dev-only Cloud KMS callout and an end-to-end /tools/execute seal mint→verify→consume + first-DENY step, and quickstart-smoke runs scripts/check_trace_conformance.py --stream over an actual live Docker Compose evidence stream.
  • Unfiltered pull_request workflow triggers for stacked PRs (POAM-2026-153) and formal-verification / audit-posture disclosures (POAM-2026-154, POAM-2026-115, POAM-2026-120) (ci(ci), ci.yml, security-scan.yml, license_guard.yml, test_workflow_pr_triggers.py, FORMAL_VERIFICATION.md, POAM.md). Drops on.pull_request.branches across all nine pull-request workflows so stacked pull requests targeting feature base branches run all seven required status checks on push (WF-M1–WF-M4), adds the explicit DistributedCBF.tla atomic-read model-vs-implementation verification-gap disclosure (POAM-2026-154), and records the 93/30-day review-bypass velocity analysis (POAM-2026-115) and test-fixture enforcement-masking pattern (POAM-2026-120).

  • CI runs terraform test against the gcp-gke target (ci, ci.yml, w6.tftest.hcl). The new terraform-test job plans the staging profile with mock providers and module overrides, so it needs no credentials and no state. Its 10 runs check what each inference backend deploys (no vLLM, advisor or NeMo server under none; vLLM and the selected advisor under vllm) and that the plan refuses external, any enable_vllm, vllm without a GPU pool, and the NeMo server without vllm. The job is not continue-on-error, so it can fail (POAM-2026-150). Two mutations of main.tf were checked to fail it.

  • CAGE_INFERENCE_BACKEND chooses the inference backend independently of jurisdiction: none, external or vllm (feat(gateway), inference_backend.py). Under none the gateway starts with no VLLM_* settings and builds no NeMo rails, and every model-dependent path refuses with its own reason code instead of admitting or reaching for a default URL.

    • Content safety answers BLOCKED with CONTENT_SAFETY_NO_INFERENCE_BACKEND (mcp_tool_server.py, semantic_rail_factory.py).
    • The inference proxy answers 503 with INFERENCE_BACKEND_NONE (inference_proxy.py).
    • Consensus at or above its threshold is a DENY with CONSENSUS_NO_INFERENCE_BACKEND, reached before any critic is consulted. Below the threshold it skips, as it does with a backend (engine.py).
    • The compliance bridge records the remediation advisory as NOT_PERFORMED with INFERENCE_BACKEND_NONE (audit_workflow.py).

    An unknown value fails startup. Unset keeps the behaviour from before the setting existed: a backend is required. No infrastructure profile changes here; choosing none per profile is a separate change.

  • A pre-merge double-check of every required status check (feat(compliance), check_pr_mergeable.sh). The script reads the required checks GitHub applies to the base branch, verifies each on the exact head SHA through the REST check-runs and commit-status APIs, fails closed on an empty rule set or any API error, and posts the result on the pull request.

  • AOF persistence with append_fsync = ALWAYS on the governance Valkey instance in every enforcing posture (POAM-2026-156 item 3) (feat(infra), memorystore_valkey, main.tf).

    • The module always renders persistence_config (DISABLED, or AOF with an explicit append_fsync) from two validated inputs, and refuses AOF without a policy and a policy without AOF.
    • The gcp-gke target names its enforcing predicate once, local.enforcing_posture (prod or staging, or any compliance flag), which the retention lock now reads too. The governance instance plans AOF + ALWAYS under it and an explicit DISABLED otherwise; the app instance plans an explicit DISABLED.
    • Applying it updates an existing instance in place: nothing under persistence_config is ForceNew in the google provider (checked at v6.43.0 and v6.50.0), and Google notes the instance might need maintenance. governance_persistence.tftest.hcl plans every committed profile. Not yet applied to staging; POAM-2026-156 closes only on the measured latency delta.
    • The enable_apac_mas_compliance description now states what the flag does: exactly what the other two compliance flags do (POAM-2026-130).

Changed

  • Evidence documents that asserted controls not in force are recorded as a class, POAM-2026-169, and five of its seven instances are corrected (docs(compliance), POAM.md, branch-protection-rules.md, TECH_STACK.md, SECURITY.md).
    • The class: a statement that a control is in force, written from a file, module or intent and never checked. Its detection rule: such a statement names the observation that showed it, with a date, or it is not evidence.
    • .github/branch-protection-rules.md now describes the protection the GitHub API returns, with the commands that read it. It had listed six settings on main that no rule applied, a release/** rule for branches that don't exist, read-and-write workflow permissions, and a release workflow that doesn't exist.
    • TECH_STACK.md no longer credits VPC Service Controls, Cloud DNS and Cloud Armor to perimeter.tf, which enables the first two in no shipped profile and attaches the third to nothing.
    • POAM-2026-115's merge counts were less than half the real ones: 233 of the 240 pull requests merged from 2026-09-30 to 2026-10-31 had no approving review, not 105. Its skew policy now records that Compliance Matrix has failed on main since 2026-10-30, and that before #602 the same failure was reported as a pass.
    • Open: the OSCAL SC-7 statements, which cite network-policy manifests no deployment path applies, and Compliance Matrix, whose workflow never installs the uv its tests need.
  • POAM-2026-115 is an accepted risk, not an open finding, and POAM-2026-120 gains a detection rule (docs(compliance), POAM.md).
    • POAM-2026-115 moves to a new Accepted Risks section. No work closes it: it closes when a second maintainer exists. Its compensating controls are stated as verified on 2026-11-02: ruleset 25966171 has no bypass actor, so --admin skips only the review rule, and check_pr_mergeable.sh --comment posts a pre-merge record on the pull request. Nothing enforces the record: 55 of the 75 merges from #552 to #632 carry one posted before the merge, and none of the 13 merged on 2026-11-01 (UTC) does. The entry had said all eight required checks were pinned to the GitHub Actions app; three were (all eight since 2026-11-02, see Security).
    • POAM-2026-120 records how to spot its next instance: a result that agrees with itself is suspect on its face.
  • Image digests and the project are rendered per deployment, never committed; deploy_all.sh builds, then renders, then plans (POAM-2026-131, re-rated High) (refactor(infra)!, variables.tf, main.tf, render_image_digests.sh, build_images.sh, deploy_all.sh, image_digests.tftest.hcl, test_image_digests_rendered.py).
    • No profile was deployable from the repository. Staging's committed digests named gcr.io/cage-reference/, which refuses every pull; every other profile inherited sha256:e3b0c442…, the SHA-256 of empty input. The renderer's output was never read, and the plan was saved before the images were built.
    • var.image_digests and var.project_id have no default, and no profile commits either. The render command is in each description (interactive prompt) and in a trailing comment on the declaration line, which Terraform quotes in its -input=false error.
    • The image_digests validation requires exactly the images the profile deploys (local.required_image_digests) and refuses the empty-input digest, so staging under inference_backend = "none" needs no vLLM, advisor or NeMo image.
    • The renderer writes a gitignored image_digests.<sha>.tfvars.json with no fallback to another build's image. deploy_all.sh renders before it plans and gains --image-sha. build_images.sh selects images with IMAGES.
    • BREAKING — infrastructure: see RD-1–RD-4 in BREAKING_CHANGES.md.
  • The GKE GPU node pool uses pool-wide total_min_node_count / total_max_node_count (location_policy = "ANY"), starts at 0 GPUs in every profile, and allows only verified regional L4 + g2-standard-8 zones (POAM-2026-132) (refactor(infra)!, main.tf, variables.tf, variables.tf, gpu_pool.tftest.hcl).
    • google_container_node_pool.gpu_nodes replaces per-zone min_node_count / max_node_count with pool-wide total_min_node_count / total_max_node_count and location_policy = "ANY", and gpu_node_pool_initial_count defaults to 0 (0 per zone is 0 total). Multi-zone sourcing no longer multiplies initial, minimum, or maximum GPU counts by length(gpu_node_locations).
    • Every .tfvars profile (dev, us-dev, eu-dev, apac-dev, staging, prod, eu-prod, apac-prod) and terraform.auto.tfvars.example sets gpu_node_pool_initial_count = 0 and gpu_node_pool_min_count = 0 (max = 2 on dev/staging, max = 5 on *prod.tfvars) and explicitly lists only verified zones offering both nvidia-l4 and g2-standard-8: us-central1-a/b/c (restoring 3-zone capacity sourcing on staging.tfvars), europe-west1-b/c (removing europe-west1-d, POAM-2026-132, closed), and asia-southeast1-a/b/c.
    • Plan-time preconditions on gpu_nodes refuse empty/out-of-region gpu_node_locations and any unverified L4/g2-* zone in those three regions, exercised against the real gcp_gke_cluster module by gpu_pool.tftest.hcl.
    • BREAKING — infrastructure: gpu_node_pool_min_count and gpu_node_pool_max_count now bound the pool total rather than each zone, gpu_node_pool_initial_count defaults to 0, *prod.tfvars ship min = 0, and europe-west1-d fails terraform plan with [POAM-2026-132]. See AG-8 in BREAKING_CHANGES.md.
  • On GKE, inference_backend decides what is deployed: staging runs the governance core with no model by default, and external is compose-only (POAM-2026-136) (refactor(infra)!, main.tf, variables.tf, staging.tfvars).
    • inference_backend replaces enable_vllm and sets CAGE_INFERENCE_BACKEND on the gateway and the compliance bridge. Under "none", the default and staging's setting, no vLLM, no advisor in either runtime and no NeMo server are deployed, so nothing is left pointing at a model service that does not exist. Under "vllm" the vLLM deployments and the advisor are deployed. "external" is refused: the cluster denies external egress by default, and opening it for the agent would weaken that control.
    • Staging keeps its GPU pool at min 0 / max 2 / initial 0, so switching to "vllm" is a variable change and an apply (W6 narrowed staging to us-central1-a while counts were per-zone; AG-8 moves the pool to total_min_node_count / total_max_node_count and restores us-central1-a/b/c).
    • enable_nemo_guardrails defaults to false. A plan guard refuses "vllm" without a GPU pool, and the NeMo server without "vllm".
    • The never-read vllm_fast_url, vllm_reasoning_url and vllm_gateway_url variables and their deploy_all.sh and load_env.sh exports are deleted (POAM-2026-136, closed).
    • Lula SI-2 requires only the deployments every backend has, and digest-checks the optional ones when present. CM-6 no longer requires nemo-guardrails-config. CM-6 still cannot pass, because nothing creates the thresholds ConfigMap it asserts (POAM-2026-151, opened).
    • The SSP no longer lists LLM provider APIs as an interconnection or an outbound inference destination on GKE. The gcp-gke README documents the backends, including that the advisor calls its models directly by design.
    • BREAKING — infrastructure: remove enable_vllm from every profile and set inference_backend. A deployment that relied on staging's vLLM, advisor or NeMo server sets inference_backend = "vllm". See AG-7 in BREAKING_CHANGES.md.
  • CAGE_NEMO_ENABLED=false now holds at runtime, not only at startup; all three opt-in switches share one parser; and "rails off" has its own reason code (POAM-2026-144) (fix(governance)!, semantic_rail_factory.py, feature_flags.py, inference_proxy.py, hybrid_server.py, mcp_tool_server.py, nemo_node_factory.py, content_safety.py).
    • Every path that builds NeMo rails consults semantic_rails_refusal(): gateway startup, the inference proxy, get_nemo_rails(), reload_nemo_rails(), the cage:nemo:reload subscriber and refinement approval. Before this change the inference proxy rebuilt rails on every request while the switch was off, and a reload broadcast re-enabled them on all sub-apps. Both routed around the nemo_rails_backend startup check.
    • With the rails off, the inference proxy runs deterministic-only: the Stage 1 keyword scan, then scrub_pii() output masking, with ISO 42001 A.6.1.2 stamped SKIPPED. The reload subscriber is not started, and applying an approved refinement returns 409 with the proposal left staged.
    • verify_content_safety returns CONTENT_SAFETY_RAILS_DISABLED when the rails are off by configuration. CONTENT_SAFETY_RAILS_UNAVAILABLE now means only "enabled but not initialised".
    • BREAKING — configuration: CAGE_NEMO_ENABLED, CAGE_CONSENSUS_CRITICS_ENABLED and CAGE_CAUSAL_TIER_ENABLED accept only true/false/1/0 (empty or unset is off). yes/no used to raise for NeMo but enable critics and the causal tier; it now raises ConfigurationError for all three. consensus_critics_enabled and CAGE_CONSENSUS_CRITICS_ENABLED_ENV move from src.gateway.governance.consensus to src.gateway.governance.feature_flags. See AG-6 in BREAKING_CHANGES.md.
  • NeMo Colang semantic rails are opt-in (CAGE_NEMO_ENABLED=false by default), validate_output_semantics() calls rails.llm.ainvoke() directly with strict SAFE parsing (POAM-2026-125), and dead CustomSelfCheckOutputAction is removed (POAM-2026-140) (refactor(nemo)!, nemo_node_factory.py, mcp_tool_server.py, hybrid_server.py, posture.py, manager.py, actions.py, action_registry.py, definitions.co, nemo-rails-configmap.yaml).
    • nemo_rails_enabled() in nemo_node_factory.py now defaults CAGE_NEMO_ENABLED to "false"; mcp_tool_server.py and hybrid_server.py skip initialize_rails() when CAGE_NEMO_ENABLED=false; and assert_production_posture() (nemo_rails_backend in posture.py) refuses startup when CAGE_NEMO_ENABLED=true and CAGE_INFERENCE_BACKEND=none.
    • BREAKING — behaviour: set CAGE_NEMO_ENABLED=true (backed by a >= 7B instruction-tuned judge model; see src/integrations/nemo/README.md) to enable Colang 2.x input flows and LLM output auditing; CustomSelfCheckOutputAction (custom_self_check_output) is removed. See AG-5 in BREAKING_CHANGES.md.
  • The STPA compiler refuses uncompilable declared enforcement targets, UCA-3/4/7/10/12 compile into structured checks, and deterministic PII redaction and Aho-Corasick screening are unbundled from get_nemo_rails() (POAM-2026-141, POAM-2026-142) (fix(governance)!, stpa_compiler.py, core_system.yaml, stpa_control_structure.yaml, trade_hazards.yaml, uca_rules.py, nemo_node_factory.py, NIST_SP800_53.json).
    • UCAModel._validate_enforcement_deps() in stpa_compiler.py now raises ValueError at parse/compile time whenever a UCA declares opa or python without a structured condition (ConditionModel.is_structured(): composite, or param + operator with threshold / threshold_ref where required) or declares nemo without condition.semantic_pattern.
    • In core_system.yaml and stpa_control_structure.yaml, UCA-7 drops its uncompiled nemo target (enforcement: [opa]). In trade_hazards.yaml, UCA-3 (output_contains_pii: is_true, enforcement: [python]), UCA-4 (partial_transaction_detected: is_true, enforcement: [langgraph, python]), UCA-10 (stale_balance_data_detected: is_true, enforcement: [python]), and UCA-12 (bounding_contract_violation_detected: is_true, enforcement: [python, opa]) now carry structured conditions and compile into src/cage_finance/stpa/uca_rules.py. NIST_SP800_53.json replaces citations to the deleted 2d4aea66^:src/gateway/governance/generated_stpa_validator.py with src/cage_finance/stpa/uca_rules.py and src/gateway/governance/stpa_validator.py, and the three OSCAL artifacts under compliance/oscal/ are regenerated.
    • nemo_node_factory.py adds nemo_rails_enabled() (CAGE_NEMO_ENABLED, raising ConfigurationError on invalid values) and runs deterministic Presidio _redact_pii() and Stage 1 ac_keyword_scan() on input, and Presidio _redact_pii() plus regex scrub_pii() on output, unconditionally before consulting get_nemo_rails().
    • BREAKING — STPA schema & validator: STPA YAML files declaring opa/python with only semantic_pattern or nemo without semantic_pattern now raise ValueError at parse time, and GeneratedSTPAValidator in src/cage_finance/stpa/uca_rules.py now enforces _check_uca_3, _check_uca_4, _check_uca_10, and _check_uca_12. See AG-4 in BREAKING_CHANGES.md.
  • Consensus separates deterministic notional threshold escalation from optional multi-model critic voting (CAGE_CONSENSUS_CRITICS_ENABLED=false by default), and proof/model.py models the 7-tier deterministic core (refactor(governance)!, engine.py, consensus_tier.py, opa_node_factory.py, posture.py, model.py). When CAGE_CONSENSUS_CRITICS_ENABLED is unset or false, ConsensusGate does not instantiate ConsensusModelRegistry or GatewayClient, and any action at or above consensus_threshold (or in high_stakes_actions) deterministically returns ESCALATE with reason_code: NOTIONAL_THRESHOLD_ESCALATION (ViolationKind.HITL → REQUIRE_APPROVAL) with zero LLM calls, including under CAGE_INFERENCE_BACKEND=none. Setting CAGE_CONSENSUS_CRITICS_ENABLED=true activates the 3-critic LLM quorum and requires an active inference backend (assert_production_posture() refuses CAGE_CONSENSUS_CRITICS_ENABLED=true with CAGE_INFERENCE_BACKEND=none). proof/model.py removes "causal" from TIERS (8 → 7 deterministic tiers: ftra, stpa, opa, confidence, consensus, cbf, fiscal; 34 gated states, 17 ungated states, 31 skipped-tier states on cbf, 38 EU_ECB states).
    • BREAKING — behaviour (staged correction to #566 493727f5): by default (CAGE_CONSENSUS_CRITICS_ENABLED=false), actions at or above consensus_threshold escalate to REQUIRE_APPROVAL (NOTIONAL_THRESHOLD_ESCALATION) instead of invoking LLM critics or failing closed with CONSENSUS_NO_INFERENCE_BACKEND under CAGE_INFERENCE_BACKEND=none; set CAGE_CONSENSUS_CRITICS_ENABLED=true to run multi-model critic voting. proof.model.TIERS now has 7 entries instead of 8. See AG-2 and AG-3 in BREAKING_CHANGES.md.
  • The finance causal tier is opt-in (CAGE_CAUSAL_TIER_ENABLED=false by default) and plugins declare required_tiers enforced at startup (refactor(governance)!, plugin.py, __init__.py, contracts.py, assembly.py, posture.py). FinanceCagePlugin and create_finance_tiers() now assemble the five deterministic finance tiers (trade_confidence, bounding, cbf, fiscal, consensus) by default and record DisabledTier("causal", "CAGE_CAUSAL_TIER_ENABLED=false") when CAGE_CAUSAL_TIER_ENABLED is unset or false. Setting CAGE_CAUSAL_TIER_ENABLED=true (or passing include_causal=True) registers CausalTierPlugin and causal_world_model_refresher; an unknown value raises ConfigurationError. PluginContribution.required_tiers declares the domain's mandatory tier names (FINANCE_REQUIRED_TIERS in finance, plus "causal" when enabled), and assert_production_posture() verifies required_tiers before any other posture check, refusing startup in enforcing postures if any required tier is absent or disabled.
    • BREAKING — behaviour: the default finance governor assembles five domain tiers instead of six; set CAGE_CAUSAL_TIER_ENABLED=true (or pass include_causal=True) to include the DoWhy/EconML causal tier. See AG-1 in BREAKING_CHANGES.md.
  • The inference proxy has no localhost fallback (fix(gateway)!, inference_proxy.py). With no endpoint configured it forwarded prompts to http://localhost:8000/v1. It now answers 503 with INFERENCE_BACKEND_URL_MISSING and rolls back the quota step.
    • BREAKING — behaviour: set VLLM_BASE_URL (or VLLM_FAST_API_BASE / VLLM_REASONING_API_BASE) explicitly. The default was undocumented, and the MCP server already refused to start without VLLM_BASE_URL. See IB-1 in BREAKING_CHANGES.md.
  • Consensus admission is decided once, in the kernel (engine.py). consensus_admits() admits only APPROVE, APPROVED or SKIPPED in status, and never a result that carries a reason_code. The finance, healthcare and physical-AI tiers use it, and record the kernel's reason code over their own.
    • BREAKING — behaviour: every domain tier refuses a consensus result without status, or one that carries a reason_code; finance used to fall back to decision when status was absent. See IB-2 in BREAKING_CHANGES.md.
  • The compliance bridge records why there is no remediation advisory (audit_workflow.py, main.py). advisor_status is NOT_REQUIRED, NOT_PERFORMED, ERROR, PERFORMED, FAILED or PENDING, with advisor_reason_code when no advisory was attempted. A required backend with no VLLM_BASE_URL is now an ERROR with INFERENCE_BACKEND_URL_MISSING, not a log line.
  • Gate G9 now covers docs/POAM.md, checks Terraform paths, and verifies removed paths pinned to a commit (check_doc_references.py). A document that has to name a path removed from HEAD pins it in git's revision syntax, <sha>^:<path>, and the gate checks the pin against history: the commit must be an ancestor of HEAD and must contain the path. It fails closed in a shallow clone, so the lint job now checks out full history. The scope lives only in the script's DOC_SCOPE; CI and make docs-check no longer carry their own lists. The label of a link with a local target is no longer checked as a second path, since the target is the reference. Three entries that only docs/POAM.md relied on left the KNOWN_DOCUMENTED_GAPS waiver. AGENTS.md records that whoever rewrites main's history owns re-pinning.

Fixed

  • Every gcp-gke profile sets enable_nist_compliance itself, so a direct terraform run with staging.tfvars plans the compliance posture (POAM-2026-174 item 1) (fix(infra)!, staging.tfvars, dev.tfvars, staging_compliance.tftest.hcl, posture_guards.tftest.hcl, outputs.tf, outputs.tf, test_control_plane_access.py, BREAKING_CHANGES.md CP-2 and CP-3, POAM.md).

    • The staging flag was commented out and only deploy_all.sh and staging_lifecycle.sh passed it. A direct terraform plan or apply -var-file=staging.tfvars therefore planned 0.0.0.0/0 on the control plane, skipped control_plane_access_guard and the POAM-019 key guards, and turned off Cloud SQL PITR.
    • Breaking for direct runs (CP-2): plan and apply now need authorized_networks (at least one CIDR, no /0) and distinct langfuse_compliance_* keys, as deploy_all.sh runs already did. Teardown does not: terraform destroy -var-file=staging.tfvars reports these resource preconditions only as warnings, verified on Terraform 1.15.8 and 1.17.1. The same check written as a variable validation would block teardown, and a comment in main.tf keeps the guards as preconditions.
    • dev.tfvars sets false, which it always got, and test_every_profile_sets_the_nist_flag_itself requires every committed profile to set the flag. The classifier in test_control_plane_access.py no longer models deploy_all.sh's default. The us-dev.tfvars comments that credited the flag with Redis replication and PDBs, and placed the key guard in app_secrets/main.tf, are corrected.
    • staging_compliance.tftest.hcl plans the profile with no flag override and asserts that compliance is active, PITR is on, only the supplied CIDRs reach the control plane, and an empty or /0 list stops the plan. A control run with the flag forced off shows the 0.0.0.0/0 fallback. Re-commenting the flag fails the file. Of the five existing test files the change touched, three supply the compliance key fixtures, so every run in them plans the new default; two force the flag off in four runs, each of which keeps the tier and flag it planned before.
    • Item 2 of POAM-2026-174 stays open: master_authorized_networks_config and database_encryption are in the cluster's ignore_changes, so a cluster created open by an earlier direct apply is not corrected by this change.
    • POAM-2026-179's remediation is amended to match. Its owner-format and TTL checks become resource preconditions, because as variable validations they would block teardown for anyone still holding a pre-convention list. Its plan-time HMAC options are marked as putting the helper on the teardown path. POAM-2026-180 is opened: the five terraform_data guards in main.tf are standalone, so a plan targeted at what one protects skips it. Two of those gaps have no preventive backstop: -target=module.gke with an empty list, and -target=module.app_secrets with shared compliance keys. POAM-2026-126, POAM-019 and the docs that describe these guards now say they refuse untargeted plans only.
    • A comment below control_plane_access_guard in main.tf holds the placement matrix (targeted plan and destroy plan, for each of four places a check can go, all run on both versions), and AGENTS.md adds the rule to its Terraform invariants: a check goes where a plan targeted at the resource it protects still evaluates it, as a precondition on that resource or a guard it depends on. The comment also records that a root-level terraform test cannot expect a failure inside a child module, so a precondition there is tested from the module's own tests. test_guard_placement.py fails on any new standalone guard in infra/. Two more false comments are corrected: main.tf said module.app_secrets skips the compliance secret when the keys are empty, and us-dev.tfvars credited the compliance bridge with detecting shared keys.
    • Breaking (CP-3): the deployment guards follow the tier, not the jurisdiction flags (POAM-2026-181). The control-plane guard, both POAM-019 key guards, the cluster allowlist, PITR and the NeMo securityContext followed the three jurisdiction flags, which all do the same thing. So a staging or prod plan that selected none planned an open control plane with no guard: the base platform with no overlay, or staging.tfvars pointed at EU_ECB with the NIST flag off as its description instructed (verified on Terraform 1.15.8 and 1.17.1). All of them now bind on local.enforcing_posture, which was already true for every staging and prod plan, and the flags decide only whether a dev plan enforces. posture_guards.tftest.hcl plans staging and prod with no flag in all three jurisdictions. test_control_plane_access.py fails if any_compliance_active gates anything else again. No committed profile is affected, and teardown is unchanged. The flag descriptions, the cluster module README, deploy_all.sh's DEP-01 comment and messages, the DEP-01 decision record and the staging row in DEPLOYMENT_RULES.md now describe the code: staging is US_FED, because deploy_all.sh ignores a CAGE_DEPLOYMENT_REGION that disagrees with the profile. OSCAL output is unaffected, because the exporter follows CAGE_DEPLOYMENT_REGION. POAM-2026-130 records that nothing checks a flag against its jurisdiction, while the jurisdiction itself is checked against the region.
  • tlc-model-check runs on every pull request and every push to main, and its runner fails on any TLC verdict the pins do not cover (POAM-2026-090, POAM-2026-091, POAM-2026-150) (ci(ci), tlc-model-check.yml, verify_tla.py, POAM.md, FORMAL_VERIFICATION.md, proof/README.md).

    • The workflow had only a workflow_dispatch trigger from 2026-10-02 and first ran on 2026-11-01 (run 41362761903). That run reproduced every pinned count, and every negative control failed as pinned. Both closures had cited the workflow as if it ran. Each POAM row now records local verification on its original closure date and the first CI run; the closure dates are unchanged.
    • verify_tla.py parsed only invariant violations, so it never read TLC's verdict on SP4_FenceEpochMonotonic, an action property declared under PROPERTIES. With SP-4 planted false, TLC v1.7.4 printed 768 Action property errors and the runner printed OK on all 8 cfgs. Under -continue TLC exits 0 and prints No error has been found even after violations. The runner now fails on any TLC Error: line other than a pinned invariant violation or its trace header, and on a run that never reports completion. Against real TLC it reports BAD on the planted tree's three DistributedCBF cfgs and OK on all 8 real cfgs.
    • test_gate_tlc_model_check_rejects_unpinned_tlc_verdicts plants six defects without Java, and the runner from before the fix fails it. tests/test_workflow_pr_triggers.py fails if either trigger is removed. POAM-2026-150 records the workflow as a sixth gate that could not fail.
    • The pins measured the state space, not what was checked over it: a cfg that dropped an INVARIANTS or PROPERTIES line kept every count and passed. DECLARED_CHECKS in tla_pins.py now pins each of the 8 cfgs' invariants and properties. The runner and test_tla_specs_proof.py both enforce it; the test also proves that dropping any one declared line is caught, and it checks the Step 15 table in FORMAL_VERIFICATION.md, including "all 11 / all 10 invariants", against the pins.
    • proof/README.md documents the out-of-repo toolchain (Java and the sha256-pinned jar). Without it, make verify-tla and pytest both stay green while TLC never runs, and CI's pytest-logic always runs that way. The gate test and the workflow note that its verdict strings are TLC v1.7.4's and must be re-derived from MP.class when the jar is bumped.
    • FORMAL_VERIFICATION.md Step 12 cited 12,184 states for the N=2 suite. That is the pre-#351 model's N=4 pin. It now cites that run's per-cfg counts, 1,811 + 3,972 + 2,388, which sum to 8,171 over three state spaces. The status row no longer says tests/test_no_direct_bind_proof.py pins 38 gated states (it pins 34), and the Step 7 output block is marked as predating #574.
  • No LLM client sits in the kernel, the base install or the gateway image (POAM-2026-168) (fix(governance)!, consensus/engine.py, seams/critic.py, critic_backend_factory.py, inference_client/, check_model_free_install.py).

    • Before this fix, assemble_governor() loaded the openai SDK for any plugin contributing consensus (finance does). It did so even under CAGE_INFERENCE_BACKEND=none, where consensus can only refuse. The base install resolved five model packages and the gateway image fourteen.
    • The consensus gate keeps its deterministic half in the kernel: trigger, context selection, prompt rendering, vote parsing and aggregation. The model call crosses the new CriticBackend seam.
    • The kernel's GatewayClient and ConsensusModelRegistry moved to Layer 3 as InferenceClient and RoutedCriticBackend. The backend is bound only when critics are enabled and a model exists. Without the inference extra, assembly fails closed with CONSENSUS_CRITIC_BACKEND_UNAVAILABLE.
    • litellm, openai and the coverage plugin left the base install, and nemoguardrails left the gateway extra. Two new extras hold them: nemo and inference.
    • The gateway image no longer installs the advisor extra or copies the advisor's source. Its comment claimed the gateway imported it; nothing does.
    • G3 now forbids LLM clients and model runtimes in src/gateway/. It listed none before.
    • The new model-free install gate runs in the Lint job. It fails if the base install or the gateway image resolves a model package, or if G3 stops listing an LLM client.
    • The hermetic job's "no ML" check covers LLM clients too.
    • A subprocess test assembles a finance governor and finds no model client in sys.modules. Its positive control sees openai once critics and a model are configured.
    • BREAKING — kernel, packaging & images: see MF-1–MF-5 in BREAKING_CHANGES.md.
  • Consequence single-use markers are replica-acknowledged on a pinned socket, provider_01 keys rec per determination rather than per mandate, and cbf_engine.py's two vacuous-WAIT defects are closed (POAM-2026-156 items 1 & 4, POAM-2026-157, POAM-2026-070, POAM-2026-090) (fix(governance)!, redis_replication.py, consequence_authority_store.py, consequence_gateway.py, cbf_engine.py, redis_single_use.py, provider.py, replicated_valkey.py, test_consequence_consumption_durability.py, test_provider_01_payload_contract.py, test_flowsignal_step2_qualification.py).

    • ConsequenceAuthorityStore.consume_once() pins a single connection via redis_replication.py (pinned_connection, write_socket, acknowledged_replicas, ReplicationPolicy.from_env()), re-asserts its own marker via _REASSERT_IF_OURS_LUA if a lost-reply retry in set_nx_outcome() took the GET branch (NxOutcome.APPLIED_UNCERTAIN), and sends WAIT directly on the underlying write_socket rather than through redis-py's retrying execute_command. Under strict replication (CAGE_STRICT_REPLICATION), an unacknowledged write raises ConsumptionNotReplicatedError without deleting the marker on the primary, and ConsequenceGateway.evaluate() returns ConsequenceDecision.BLOCK with reason_code="CONSUMPTION_NOT_REPLICATED".
    • provider_01 removes the test-only prefer_receipt_id_for_rec payload flag, unconditionally sets ConsequenceToken.rec to authority_receipt.id (or authority_receipt_id), and fails closed (cage.authority_denial) when an ALLOW response carries no receipt id. TC-01–TC-06 in test_flowsignal_step2_qualification.py now build their payload through FriaTier._payload(), and test_provider_01_payload_contract.py guards against top-level payload reads outside FriaTier._payload().
    • Extracting redis_replication.py also remediates two distinct defects in cbf_engine.py that left its prior WAIT vacuous against DistributedCBF.tla's SyncReplication: (1) rollback_state() passed the unpinned pool client (self.redis_client) into _sync_to_replicas() so LUA_ROLLBACK_CBF and WAIT could run on different pooled connections on every call with no disconnect involved, and (2) atomic_verify_and_commit() / _sync_to_replicas() called target_client.execute_command("WAIT", ...) through the retrying client so a socket drop during primary failover silently reconnected on a fresh socket with zero writes and returned >= 1 vacuously. Nine mutations (M1–M9) are verified in POAM.md.
    • BREAKING — governance & provider_01: see CD-1–CD-3 in BREAKING_CHANGES.md.
  • Declared consequence magnitudes (ConsequenceMagnitudes) are enforced at FriaTier pre-flight, ConsequenceToken minting, and ConsequenceGateway consumption (POAM-2026-155) (fix(governance)!, contracts.py, consequence_magnitude.py, fria_tier.py, consequence_token_service.py, consequence_gateway.py, assembly.py, provider.py, adapter.py, src/cage_finance/, src/cage_healthcare/).

    • Domain plugins declare per-action ConsequenceMagnitude tuples over MagnitudePath(keys, minor_unit_digits) (paths=() when explicitly magnitude-free). ConsequenceMagnitudes.violations() refuses undeclared actions (MAGNITUDE_UNDECLARED), absent paths under params (MAGNITUDE_ABSENT), and values that are bool, non-numeric, non-finite, sub-minor-unit dust, <= 0, or > 2^53 - 1 minor units (MAGNITUDE_INVALID).
    • Enforced before provider.validate_fria() in FriaTier.evaluate() (FRIA_MAGNITUDE_INVALID), before signing in mint_consequence_token_finding() (CONSEQUENCE_TOKEN_MINT_FAILED), and at Step 5 of ConsequenceGateway.evaluate() before store.consume_once() (leaving the single-use authority record unconsumed on refusal). provider_01 drops its name-based "amount"/"magnitude" float checks.
    • cage_finance declares all 6 registered actions (execute_trade, execute_trade_bounded at minor_unit_digits=2; 4 read-only actions with paths=()); cage_healthcare declares its 5 dosing actions at minor_unit_digits=6 (whole nanograms) and leaves adjust_dosage and override_contraindication undeclared (MAGNITUDE_UNDECLARED under EU_ECB); cage_physical_ai remains unrunnable (domain_config = None, POAM-2026-077) with consequence_magnitudes=().
    • BREAKING — governance & integrations: magnitudes is a required keyword-only argument on mint_consequence_token_finding() and ConsequenceGateway(), minor_unit_digits is required on MagnitudePath, and undeclared actions fail closed with MAGNITUDE_UNDECLARED. See CM-1–CM-4 in BREAKING_CHANGES.md.
  • The GKE load test restores the gateway's replica count when its stdout closes or its terminal hangs up (fix(infra), run_gke_load_test.sh, test_gke_load_test_script.py).

    • The script pins the gateway to one replica and restores the original count from its EXIT trap. It trapped INT and TERM but not SIGPIPE. With stdout closed (a | tee reader gone, a dropped session), its next write raised SIGPIPE, which killed bash before the EXIT trap could run, so the gateway stayed at one replica. Against the test stubs this happened every time stdout closed with no signal (returncode −13, no restore call). It also happened when SIGTERM arrived after the script's next write, which is how test_closed_stdout_pipe_still_restores_the_original_count failed once under parallel load.
    • SIGPIPE and SIGHUP are now trapped like INT and TERM, and all four exit through the restore. Bash already ran the EXIT trap for an untrapped SIGHUP, but only through its own signal handling.
    • New tests close stdout with no signal (exit 141) and send SIGHUP (exit 129). The close-then-SIGTERM test accepts either signal handling first (141 or 143); the restore is asserted in every case.
  • The Langfuse event bucket shares the WORM bucket's retention posture, so a staging walk locks it for one day instead of seven years (POAM-2026-159) (fix(infra)!, main.tf, outputs.tf, retention_posture.tftest.hcl, test_clickhouse_operator.py, DEPLOYMENT_RULES.md).

    • local.retention_period_seconds (86400 s for staging, 220752000 s for every other tier, the dev default included) and local.retention_locked (always for staging and prod, for dev only under a compliance flag) feed both google_storage_bucket.langfuse_event_upload and module.worm_bucket. The Langfuse bucket locked 2555 days whenever a compliance flag was set, which deploy_all.sh does for US_FED staging, so every staging walk left a bucket, and a lien on its project, that nobody could remove for seven years.
    • GCS refuses to shorten a locked policy, so the bucket moves to google_storage_bucket.langfuse_event_upload (langfuse-event-upload-<project>-<suffix>), and a removed block (destroy = false) makes the first apply forget google_storage_bucket.langfuse_events without deleting it. On the live staging project the old bucket and its lien stay until at least 2033-09-27. DEPLOYMENT_RULES.md gives the migration, why it must be applied before any destroy, and what still stops a full staging teardown: signing keys with prevent_destroy, buckets without force_destroy, and the project lien.
    • Twelve plan-only runs on mock providers read every committed profile from its file: staging takes one day on both buckets, every prod profile keeps 2555 days, and an unknown tier is refused. Two static checks pin that both buckets read the shared locals and that an unset tier defaults to dev. Fourteen mutations of the gate (M1–M14 in the pull request) each fail a named test.
    • The IAM comment no longer calls the bucket a transient queue.
    • BREAKING — infrastructure: see LR-1 and LR-2 in BREAKING_CHANGES.md.
  • The POAM-019 plan guard is split into one resource per condition, and each error headlines the condition it detected (fix(infra), main.tf, poam_019.tftest.hcl, DEPLOYMENT_RULES.md, QUICK_START.md).

    • With a compliance posture active, terraform_data.poam_019_empty_key_guard refuses empty compliance telemetry keys ([POAM-019] Compliance telemetry keys are empty under compliance hardening.) and terraform_data.poam_019_shared_key_guard refuses keys equal to the application project's ([POAM-019] Compliance telemetry keys equal the application project keys.). The single guard they replace headlined both as an AU-9 isolation breach, so unset variables ("" == "") read as a compliance defect when .env had not been loaded. An empty key is no longer also reported as shared, the likely cause moves into the remediation text, and both messages cite docs/compliance/us_fed/POAM_US_FED.md instead of the nonexistent docs/POAM_US_FED.md#POAM-019.
    • The POAM-2026-086 Memorystore CA precondition that shared the old resource moves unchanged to terraform_data.memorystore_ca_pinning_guard. Existing state plans one destroy (the old resource) and three creates; none of them manages infrastructure.
    • Seven plan-only runs on mock providers test each guard on its own, since expect_failures names a resource, not a precondition.
    • Troubleshooting entries carried over from the closed PR #591: the Linkerd network-validator crash loop (root cause unestablished), the unverified sa-cage-deployer path, the POAM-2026-126 authorized-networks lockout, the POAM-019 diagnostics, and diagnostic hygiene.
  • The partner over-the-wire harness can fail, and runs every partner with only its own secrets (POAM-2026-150) (ci(ci), ci-partner-integration.yml, partner_harness.py, require_live.py, selection_flags.py, test_partner_harness.py). It had never run, and as written no run could have tested a partner.

    • --run-partner-integration now enables partner tests. Every partner live test also carries live_external, which conftest.py skipped unless --run-live-external was set, so the harness, make test-partner and the runbook's commands all skipped every partner test.
    • One job per partner (provider_01, 02, 03, 05, 07, 08, actuator_01; 07 and 08 were missing), at the paths #218 moved the tests to. Each secret reaches only its own partner's test step; before, the secrets sat on a step that ran no tests.
    • The require_live plugin fails a job on any skip, and the verdict job fails unless every partner job passed. Before, an all-skipped run and a skipped job both reported success.
    • actuator_01 receives the file paths its adapter reads: its PEM secrets are written to files at run time. The workflow used to pass differently named variables that the adapter never read.
    • Runs weekly as well as on dispatch. test_partner_harness.py checks the workflow against the live tests on every pull request.
  • Live quickstart-smoke exercises Theorem 5.3 routing-seal nonce replay (INV-2) and HTTP transaction_id replay against a running Redis container, and POAM-2026-115 / branch-protection docs match live ruleset 25966171 (eight required status checks) (ci(ci), ci.yml, test_quickstart_bootstrap_contract.py, test_workflow_pr_triggers.py, branch-protection-rules.md, POAM.md).

    • quickstart-smoke now replays (1) the already-executed AAPL transaction_id over POST /tools/execute (ExecutionLedger Redis SET NX emits GOVERNANCE_REFUSAL_RECEIPT and refuses a second actuation) and (2) a live EdDSA JWT routing seal through verify_and_consume_seal() twice inside the running gateway container against redis:7.2-alpine (SymbolicGovernorViolation: replay attack detected — seal already consumed), asserting ACTUATION_RECEIPT == 1, GOVERNANCE_SEAL_CONSUMED == 2, and --expected-refusal-count 2 via check_trace_conformance.py --stream.
    • POAM.md (POAM-2026-115, POAM-2026-153), .github/branch-protection-rules.md, and test_workflow_pr_triggers.py now document all eight required status checks in ruleset 25966171 (including Quickstart Bootstrap & First-DENY Smoke Gate, integration_id: 15368).
  • Concurrent CBF commits no longer false-regress the fence epoch (POAM-2026-154), and make lint runs mypy src/ (POAM-2026-153) (cbf_engine.py, test_cbf_settlement_ledger.py, Makefile).

    • DistributedCBF.tla models Read as an atomic observation of available_balance, fence_epoch, and agent_epochs[a], whereas cbf_engine.py split it across two multi-await windows: _get_fence_epoch() + _check_fence_epoch() and a second post-read HWM check in atomic_verify_and_commit(). When a concurrent commit landed inside either window, the losing coroutine logged CRITICAL CBF_EPOCH_REGRESSION_DETECTED or FENCE_EPOCH_REGRESSION and aborted on attempt=1 instead of retrying — both flaking the POAM-2026-070 double-spend guard (test_concurrent_trades_over_the_limit_admit_one[one-cbf]) and rejecting concurrent admissible trades.
    • _read_fence_epoch_and_hwm() now reads (safety:fence_epoch, safety:fence_epoch_hwm) in one MGET (and in one transaction pipeline in self-reported mode), _read_cbf_state_atomic() brackets the read with seen_before / epoch_before to re-read on forward concurrent commits while failing closed immediately on genuine stale-replica regressions, and atomic_verify_and_commit() reuses the pre-read verified epoch and atomically captured HWM. Deterministic skew tests (skew_ticks = 0..13) and mutations CBF-M1–CBF-M4 guard every path.
    • Makefile's lint target now runs uv run mypy src/, closing the local-gate omission exposed when stacked pull requests #578–#582 skipped main-filtered CI workflows (POAM-2026-153, opened for the workflow branch filters).
  • The quickstart and README describe the core as merged (QUICKSTART.md, README.md, proof/README.md).

    • The quickstart expected /health to report "nemo":"active". NeMo rails have been opt-in since W5, so under the compose defaults it reports "off". test_quickstart_bootstrap_contract.py now compares the documented response with the gateway's, so the two cannot drift apart again unnoticed.
    • The README's pipeline ended → consensus → causal. It now shows the deterministic core in the formal model's order, ending → consensus → CBF → fiscal, and says that consensus escalates to a human, critics are optional and the causal tier is an optional extension.
    • proof/README.md records that the downscoping plan predicted 34 / 19 / 32 states with FULL / POST_HITL / DRY_RUN labels, and that the measured 34 / 17 / 31 / 38 are the values the tests pin.
  • The chaos-agent playground no longer presents emulations as real components, and its trades can be allowed (chaos_agent_playground.py, examples/README.md).

    • Only Tier 0, the finance plugin's GeneratedSTPAValidator, is real. Yet Tier 2 printed a Python re-implementation of the STPA Rego rules under "OPA Rego Policy", and the README said every layer "executes its actual logic". Every tier now says whether it is real, emulated or simulated, and the README and QUICKSTART say the playground never calls the governor, OPA, consensus, CBF or a model.
    • Every trade payload lacked order_size and daily_vol, so UCA-6 refused every trade whatever the other inputs. The payloads now carry them. A new Scenario F, a compliant baseline, shows the real validator allowing the action it denies in Scenario A, and the summary checks each scenario against its expected verdict.
  • deploy_all.sh takes the posture from the profile, not from the --env name (deploy_all.sh; ADR-005 in DEPLOYMENT_DECISION_RECORD.md). The banner and enable_nist_compliance followed the --env name and CAGE_DEPLOYMENT_REGION, while terraform applied the profile's environment and cage_deployment_region. --env staging or --env prod without CAGE_DEPLOYMENT_REGION skipped the NIST hardening their US_FED profiles declare, and --env eu-prod printed "DEV MODE". A CAGE_DEPLOYMENT_REGION that disagrees with the profile is now ignored with a warning. Two error messages no longer tell operators to put Langfuse compliance keys in a committed tfvars file.

  • Twelve references in docs/POAM.md and one in the PIA pointed at nothing, or at the wrong file (POAM.md, PRIVACY_IMPACT_ASSESSMENT.md). Seven name files whose deletion was the remediation or that were removed later; they are pinned to the commit before the deletion rather than repointed. One of them is POAM-2026-032's scaffold, which a documentation sweep on 2026-09-16 (a2a338b8) had repointed to a live production module that the entry does not describe. Three cited files never existed: POAM-2026-069's and POAM-2026-070's test module and POAM-2026-086's Terraform module. One test had moved (#218), and one verification step is reworded. The PIA cited an IAM file that has not existed at any commit; it now cites infra/targets/gcp-gke/iam.tf.

  • The compliance artifacts no longer credit NeMo with controls the deterministic path provides (POAM-2026-145) (fix(compliance), oscal_ssp_exporter.py, NIST_SP800_53.json, aarm_mapper.py, compliance/oscal/). NeMo rails have been opt-in since AG-5 (#576), but the artifacts still described them as the enforcement.

    • The SSP listed the NeMo Guardrails Server as operational for ISO 42001 A.9.2 and NIST PT-2/PT-3. Those claims now belong to a new Deterministic PII Redaction component: in-process Presidio (12 identity entity types, not 16) and regex scrub_pii. Wherever its statements claim fail-closed behaviour they carry the input-side fail-open recorded below as POAM-2026-147, and they stay that way while that item is open. The NeMo component is other, an optional tier. The system description, data flow, boundary and PII information type say the same.
    • The STPA narratives cited config/rails/generated_stpa_rails.co, which defines no rails, as enforcing nine SP 800-53 controls (eleven mentions in the SSP with the artifact list and the artifact-nemo prop). The exporter now names only the generators a control's UCAs target, so the rails file, the prop and "NeMo Colang" in the compiler component appear only while some UCA targets nemo. The narratives also stop citing OPA for controls whose UCAs target only Python.
    • The FTRA SI-10 statements credited "NeMo/AGP rails generated from UCA-7". UCA-7 compiles only to OPA; the always-on injection screen is the Stage 1 keyword scan. The SP 800-53 component that claimed SI-10 and SI-3 through NeMo, at a path that does not exist, is now the deterministic screening component, and SI-3 is partial because nothing deterministic screens model output for harmful content. AARM-V2, V5 and V10 cite the deterministic code, and V10 no longer claims 20 entity types.
    • POAM-2026-094 and POAM-2026-114 are re-scoped to defects of the optional causal tier (CAGE_CAUSAL_TIER_ENABLED, off since #573). Three findings from the review are recorded as open: A.9.2's Lula evidence has no producer (POAM-2026-146), input redaction fails open when Presidio cannot initialise (POAM-2026-147), and the component definition describes a TypeScript gateway that never existed in this repository (POAM-2026-148).
  • Test sessions no longer read .env, and a hermetic session refuses to start while one can reach it (POAM-2026-120) (fix(tests), dotenv_guard.py, conftest.py, test_dotenv_guard.py, ci.yml).

    • tests/conftest.py loaded the repository's .env in three places, one of them a bare load_dotenv() that also searched every parent directory. git clean keeps the file, so one left in a reused worktree changed every later run: on 2026-11-01, one written by a quickstart run turned 87 tests red. The same file could as easily have turned failing tests green.
    • The loaders are gone, but litellm, which src imports, still calls load_dotenv() at import, and the virtualenv's place inside the repository puts the repository's .env on its search path. So the session checks instead of trusting: if a .env sits in the test directory, the working directory, site-packages or any of their parents, a hermetic session stops with a usage error that names the file.
    • Live runs (--run-integration, --run-live-external, --run-partner-integration, --run-e2e) are exempt. They take their values from the shell, or explicitly from uv run --env-file .env. CI's quickstart job runs its pytest step before it creates .env for Docker Compose.
    • Contributors: a hermetic run beside the quickstart's .env now stops, where QUICKSTART §6 previously only asked you to move the file aside.

Security

  • main can no longer be force-pushed or deleted, all eight required checks are pinned to GitHub Actions, and GitHub Actions can no longer approve pull requests (POAM-2026-170, POAM-2026-115) (repository settings; POAM.md, branch-protection-rules.md, SECURITY.md).

    • No rule on main blocked a force push or deletion: classic protection was off, and its rulesets held only the review rule, which admins bypass, and the required checks. The activity log records two force pushes to main, both on 2026-07-09. Ruleset 26377312 now applies non_fast_forward and deletion to main with no bypass actor. It was seen refusing a force push and a deletion on a disposable branch before it was retargeted to main.
    • Five of the eight required checks in ruleset 25966171 accepted a result of the same name from any source. All eight are now pinned to the GitHub Actions app, the only source any of them had used.
    • A workflow's token could approve a pull request and so meet the review rule without anyone reviewing. That setting is off; no workflow used it.
  • Mesh identity is accepted only from the pod's own Linkerd proxy, and governance pods without the proxy are refused at admission (POAM-2026-167, High) (fix(security)!, workload_identity.py, auth.py, proxy-required-policy.yaml, test_workload_identity_middleware.py, test_mesh_proxy_required_policy.py).

    • The gateway and the compliance bridge took the caller's identity from l5d-client-id without checking where it came from. Through the proxy the header is sound: on staging, the proxy replaced a forged value with the caller's certificate identity and removed an unmeshed caller's. A pod without its proxy took whatever a caller sent. That included the approver identities on the gateway's approval routes and the operator identity on the bridge's dual-control path.
    • Pods without the proxy do get created. The injector's webhook fails open, and on 2026-10-29 a gateway replica was admitted without the proxy while the injector was stuck.
    • The header now counts only on a connection from inside the pod (is_pod_local_connection): loopback, or the pod's own IP, which is where the proxy connects from. From any other peer, the gateway refuses the request (INGRESS_IDENTITY_NOT_FROM_PROXY) and the bridge returns 401. Both, and the standalone MCP server, run uvicorn with proxy headers off, so X-Forwarded-For cannot set the peer.
    • The service_mesh module's proxy_required_namespaces installs the cage-mesh-proxy-required ValidatingAdmissionPolicy. A pod created in a listed namespace without the linkerd-proxy container is refused, with no opt-out, and its controller retries until the injector is back. The gcp-gke target lists its governance namespace. Tested live in a staging probe namespace.
    • BREAKING: a request that sets l5d-client-id itself, from outside the pod, is refused. Only kubectl port-forward, which connects from inside the pod, still works. Docker Compose has no proxy, so compose_gateway_call.sh sends identified requests from inside the gateway container; the quickstart and its CI gate use it. On gcp-gke, a governance-stack pod can no longer opt out of the proxy.
  • The pip-audit gate's four suppressions are removed, since none of them matched a locked package version, and an --ignore-vuln without a comment naming its ID now fails the tests (POAM-2026-161) (ci(security), security-scan.yml, test_workflow_vuln_suppressions.py).

    • PYSEC-2026-139 and GHSA-rrmf-rvhw-rf47 covered torch, which left uv.lock in #224; #90 had deleted the second one's comment and kept the flag. PYSEC-2026-3721 is fixed in pip 26.2, the version the lock pins and the audit venv installs. GHSA-rg7c-g689-fr3x was commented as the advisory for diskcache, which is not in the lock either, but the ID is google-adk's CVE-2026-4810, and the locked google-adk 2.3.0 is not affected. POAM-017 in POAM_US_FED.md, which still said CVE-2026-4810 was ignored in CI through that ID, now records the removal.
    • The "Scan compliance extras" step audited the same venv as the gate, which already holds every extra including compliance, and ended in || true, so it could not fail. It is deleted.
    • Under the runner's bash -e, the gate step exited before printing its JSON whenever pip-audit failed. It now prints the results, then fails.
    • The test parses every workflow. A flag passes only if a shell comment in its own run: block names the exact ID; flags outside a run: block, non-literal IDs and the abbreviations argparse accepts in a pip-audit command, such as --ig=, are reported.
  • The gateway's egress allowlist no longer names model providers (POAM-2026-152) (variables.tf, egress-lockdown.yaml).

    • api.openai.com, api.anthropic.com and generativelanguage.googleapis.com are removed from the Terraform default and from the static manifest the SSP cites as SC-7 evidence. No GKE inference path called them: inference is in-cluster vLLM or none, and the external backend is compose-only. A test keeps the two lists equal and free of providers.
    • Still open: the *.googleapis.com wildcard the gateway needs for GCP APIs also matches Google's model API. Narrowing it needs an inventory of the GCP APIs the gateway calls.
    • The compliance documents that still described third-party providers as part of the system are corrected: the SP 800-53 profile's PT-2 guidance (which also repeated "16 entity types" and credited NeMo; it is 12, deterministic Presidio), the catalog's SA-9 and the SSP outline.
    • POAM-2026-149 records the staging WORM check: none of the 30 archived actuation receipts holds a seal, because the evidence sink's sanitizer redacted each one as [REDACTED_JWS]. That prevented the leak (no irremovable credential) and permanently broke decision linkage for those 30 receipts; restoring linkage must not weaken the sanitizer's coverage.
    • Healthcare no longer declares a causal graph that nothing read (POAM-2026-143, closed).
  • langgraph patched for GHSA-3fx2-cqw9-xr3c (Dependabot #54–#56, moderate) (uv.lock, packages/cage-client/uv.lock). langgraph 1.2.7 → 1.2.14 in the root lock; langgraph 1.2.11 → 1.2.14 and langgraph-sdk 0.4.4 → 0.4.6 in the cage-client lock. Both are runtime dependencies. Nothing in src/ or packages/cage-client/ imports langgraph_sdk or RemoteGraph, so the vulnerable streaming client was not reached; patched because a fix exists.

  • CI jobs that cannot fail are recorded as a pattern (POAM-2026-150) (POAM.md). ci.yml's integration-smoke job runs pytest -m integration without --run-integration, so it skips all 153 tests it selects and passes. It has never run an integration test since it was added on 2026-08-11. It is the fourth job found this week whose test outcomes cannot fail the build, after regional-posture.yml, the test step of compliance-matrix.yml (both POAM-2026-121) and lula-validate.yml. In its run on 2026-10-29, lula-validate.yml satisfied none of the 33 manifests and succeeded. Every Lula result committed to the repository is a schema error. The entry records the pattern and changes no workflow. While it is open, none of these jobs' status is cited as evidence that something was tested.

  • Every writer of a decision digest validates it, and each digest field has one meaning (POAM-2026-149, reopened) (fix(security)!, digests.py, quarantine.py, verdicts.py, quarantine_actuator.py, ocsf_ingestor.py, stream_invariants.py). The 2026-10-29 fix below covered one of the three writers of governance_decision_digest. QuarantineDirective had no format check and two meanings: the kernel's quarantine set it to a refusal receipt's proof_hash, and actuator_03 to a clearance's seal digest. actuator_02's OCSF model accepted any string of 8 to 128 characters. Both wrote the field into the WORM-archived evidence stream, and the oracle checked it only on actuation receipts.

    • One validator, require_sha256_hex() in a new seam module that imports only the standard library, accepts only lowercase SHA-256 hex matched in full, and its error never echoes the value. ExecutionClearance, QuarantineDirective and OcsfEventModel use it; the OCSF model also checks digest keys nested in metadata and hides inputs in its error strings.
    • governance_decision_digest now means only seal_digest(seal). A refusal-triggered quarantine names its refusal in a new field, refusal_proof_hash. A directive carries exactly one of the two, to_dict() emits only that one, and WORKLOAD_QUARANTINE_RECEIPT records it under its own name. Should a refusal hash ever be malformed, handle_deny() refuses the quarantine and still raises the DENY.
    • The stream oracle gains digest_format: both fields, at any depth of every record of every event type, must be lowercase 64-hex, and a finding never echoes the value. A refusal-triggered quarantine receipt must follow its refusal (refusal_evidence), and _is_hex64 now matches in full. The packaged copy is byte-identical, and cage-conformance 0.3.0 lists the new rule.
    • The fixtures built around the defect ("digest-abc-123", "digest-1") now carry real digests. test_decision_digest_format.py covers every writer, the oracle and the packaged copy; in a recorded mutation check, each of DD-M1–DD-M26 fails a named test. POAM-2026-149 satisfies both technical closure criteria (DD-M1–DD-M26 and cage-conformance 0.3.0) and awaits only the merge SHA; over-the-wire partner sandbox verification of actuator_03's changed signed directive (partner notice) is tracked separately in POAM-2026-166.
    • BREAKING — contract and wire: QuarantineDirective is keyword-only and takes exactly one of governance_decision_digest and refusal_proof_hash; a kernel quarantine's receipt and actuator_03's signed body carry refusal_proof_hash; actuator_02 refuses a non-digest; stream verification reports digest_format. See DV-1–DV-4 in BREAKING_CHANGES.md.
  • Actuation evidence records the routing seal's digest, never the seal (POAM-2026-149) (fix(security)!, actuation.py, routing_seal.py, broker_actuator.py, finance and healthcare tool_provider.py). Finance from #190 and healthcare from #534 passed the raw routing seal as ExecutionClearance.governance_decision_digest, and from #327 dispatch_actuation() wrote that field into every actuation receipt on the evidence stream. The stream's durable copy is WORM-retained, so a seal written there can never be redacted. Each seal had been consumed before its receipt was written and expired after the seal TTL (30 s by default), so the current verifier refuses a replay; the problem is that the archived copy is permanent.

    • Both domains now record seal_digest(seal), the SHA-256 the decision trace already uses as seal_ref, so a seal's issuance and its actuation join on one key. ExecutionClearance refuses any digest that is not lowercase SHA-256 hex, so a seal in that field fails at construction in any domain. The finance broker takes the seal it re-checks from the out-of-band routing_seal, not the digest.
    • test_seal_digest_evidence.py mints real HMAC and JWT seals and follows both domains through dispatch_actuation() to the evidence record; in a recorded mutation check, each of SD-M1–SD-M13 fails a named test.
    • BREAKING — contract: an ExecutionClearance whose governance_decision_digest is not a lowercase SHA-256 hex digest raises ValueError, and the broker's refusal for a missing seal is MISSING_ROUTING_SEAL (was MISSING_GOVERNANCE_DIGEST). See RS-1 in BREAKING_CHANGES.md.
  • A consensus critic with no endpoint no longer falls back to api.openai.com (POAM-2026-127) (fix(gateway)!, llm.py). With only VLLM_BASE_URL and VLLM_API_KEY set, the documented minimum for CAGE_INFERENCE_BACKEND=external, a critic without a dedicated URL built an OpenAI client with no base URL. That client targets OpenAI's public API with the operator's key and the trade prompt. Critics now use the routed endpoint, else VLLM_BASE_URL, else refuse, and a refusal is an ERROR vote.

    • BREAKING — behaviour: with no endpoint configured, critics vote ERROR and consensus escalates instead of calling OpenAI. See IB-4 in BREAKING_CHANGES.md.
  • Staging's GKE control plane no longer accepts every address (POAM-2026-126) (fix(infra)!, main.tf, staging.tfvars). staging.tfvars authorized 0.0.0.0/0 under a header claiming VPN scope, and nothing read the list. Whenever a compliance posture is active, the plan now refuses an empty allowlist or a /0. The live cage-staging allowlist was narrowed the same day.

    • BREAKING — deploy contract: supply authorized_networks (TF_VAR_authorized_networks or a --var-file after the profile) when deploying staging or any profile with a compliance flag on. See CP-1 in BREAKING_CHANGES.md.
  • A content-safety check that reaches no verdict is now a refusal (POAM-2026-124) (fix(gateway), manager.py, mcp_tool_server.py, api.py). With no NeMo rails, the gateway's verify_content_safety answered an error string that the advisor labelled SAFE. When NeMo raised "No main flow found", three rail functions returned safe. Each such path now returns BLOCKED with CONTENT_SAFETY_RAILS_UNAVAILABLE, CONTENT_SAFETY_NO_MAIN_FLOW or CONTENT_SAFETY_ERROR (content_safety.py), keeping the degraded span marking, and the advisor records only an exact SAFE as safe. test_content_safety_no_verdict.py pins each path; in a recorded mutation check, each test fails when its path is reverted.

    • BREAKING — output: absent rails answer BLOCKED: CONTENT_SAFETY_RAILS_UNAVAILABLE: … instead of ERROR: NeMo rails not initialised., a check that raises refuses instead of failing the tool call, and refusal reasons lead with a reason code. See IB-3 in BREAKING_CHANGES.md.
  • Output semantic validation calls rails.llm.ainvoke() directly and accepts only an exact SAFE verdict, and dead CustomSelfCheckOutputAction is removed (POAM-2026-125, POAM-2026-140) (refactor(nemo)!, manager.py, actions.py, test_validate_output_semantics.py). validate_output_semantics() previously called rails.generate_async() with a system message, routing through the Colang input flow without ever sending the auditor prompt (POAM-2026-125). It now rejects empty output_text, calls rails.llm.ainvoke() directly, strips closed <think>...</think> blocks, accepts only exact ASCII "SAFE", extracts the reason on "UNSAFE", and fails closed with CONTENT_SAFETY_ERROR on all other replies. Dead CustomSelfCheckOutputAction (custom_self_check_output in config/rails/actions.py), which no Colang flow invoked and whose Stage 2 safe_indicators short-circuited on 22 common financial words, is removed (POAM-2026-140). Guarded by test_validate_output_semantics.py (OUT-M1–OUT-M6 mutation table verified in POAM.md).

  • POAM-2026-070 was closed while its control did not operate; a regression test now guards it, and POAM-2026-119 and POAM-2026-120 are opened (POAM.md, test_cbf_settlement_ledger.py). POAM-2026-070, an externally reported concurrent double-spend within the reconciliation window, was closed on 2026-08-28 citing a function and a test module that have never existed. The debit tracking #102 added was not atomic. On Redis 7.4.12, two trades each admissible alone but not together both committed when run concurrently, at 21ad845a and at 46447c2f, the parent of #256; run one after the other, the second was refused. From #256 (2026-09-26) the fence-epoch compare-and-set refused the second concurrent commit; since #506 (2026-10-22) it retries, and the in-script debit total refuses it. The entry now records both dates: the original closure (struck, incorrect), the control operating from 2026-09-26, and verification on 2026-10-27.

    • test_concurrent_trades_over_the_limit_admit_one runs the two trades concurrently on one CBF instance and on two, with every Redis call yielding. In a sensitivity check it fails when the debit write is moved out of the Lua script, as #102 had it, even with today's fence-epoch check in place, and passes when only the fence-epoch check is removed.
    • POAM-2026-023 now records that its remediation removed the in-script balance read that had serialized concurrent commits. POAM-2026-069's test pointer and refusal message are corrected. The POA&M/Lula drift checker had counted all three as covered by Lula files that match their control IDs and never mention the CBF; it now counts them by their tests.
    • POAM-2026-119: of 13 identifiers cited as commits, six are pre-squash branch commits, each mapped to the squash commit that carries it, and seven resolve nowhere, including the only commit POAM-2026-011's closure cites. None is an advisory-fork commit kept in the maintainer's private refs.
    • POAM-2026-120: closed controls are not guarded by tests that fail when the control stops operating. 34 of the 80 closed entries cite no test, and the drift checker's 52 "covered" entries rest on an unchecked file-name match. The remediation is a forward closure requirement and a backfill of High entries only.
  • Required status checks on main can no longer be bypassed, and release tags can no longer be moved or deleted (repository rulesets; POAM-2026-115 and POAM-2026-116 in POAM.md). The checks moved out of the review ruleset into ruleset 25966171, which has no bypass actor. Before that, the maintainer's admin merges bypassed both rules, and 163 of the 215 pushes to main from 2026-09-27 to 2026-10-27 went in with required checks not passing. The one exception is an advisory fix landed from a private fork, where Actions does not run: the ruleset is disabled for that merge and re-enabled, as SECURITY.md describes. Tag ruleset 25965917 protects every v* and client-v* tag, including v4.0.0, with no bypass actor; no tag was protected before. POAM-2026-115 now covers only AC-5, since one maintainer cannot approve their own pull requests. .github/branch-protection-rules.md no longer claims that code-owner review is enabled, and its tag section describes the new ruleset.

  • The causal risk boundary can fail open as volatility rises; POAM-2026-117 records it and POAM-2026-094 may not close first (POAM.md; documentation only). The kernel predicts risk as 0.5 + β × amount / normalization_scale. That prediction has no volatility term and uses a 0.5 that is neither configured nor estimated. In the frame POAM-2026-117 states, the boundary permits $25,504.81 at volatility 0.9, while a volatility-aware reading of the same fit trips at $9,619.17. The defect is latent while POAM-2026-094 leaves every causal verdict untrusted in enforcing postures. POAM-2026-094 is now amended to say it must not close before POAM-2026-117, and its scope now covers the kernel's MockTelemetryProvider, which generates the same calibration formula as the finance fixture. POAM-2026-118 records that staging and local can select the mock telemetry provider: only prod and production are refused, and only the mock's missing timestamp column stops it today. Closed POAM-2026-057, which records the formula as implemented, now cross-references POAM-2026-117.

  • §3 of the Regional Posture Validation Report makes a false composition claim, a plugin overlay can redefine a kernel control, and content-safety checks that did not run are reported as passed; POAM-2026-121 to POAM-2026-124 record them (POAM.md, REGIONAL_POSTURE_VALIDATION_REPORT.md; documentation only).

    • POAM-2026-121 (High): §3's independence claim was false when written, because cage_finance's US_FED overlay redefines the kernel's CTRL_FTRA_001. Its 12 hashes reproduce exactly on 3b69207c, the parent of the commit that published them, and match no tree that contains the report. §3 and the report's head now carry notes saying not to rely on it. The claim is withdrawn rather than proven: POAM-2026-123 shows independence violated in a path that changes a tier's verdict.
    • POAM-2026-122 (Low): region-pinned tests number 57 for eu_ecb, 23 for us_fed and 13 for apac_mas, although US_FED carries the NIST SP 800-53 surface.
    • POAM-2026-123 (High): the control registry applies plugin overlays over kernel controls and only logs the collision. A test overlay that defines CTRL_FRIA_006 moved the EU_ECB FRIA tier past its Art. 27(2) staleness gate. With the kernel baseline alone, that gate denies without consulting the provider.
    • POAM-2026-124 (High): with no rails, or when NeMo has no main flow, verify_content_safety and the advisor report a check that did not run as passed. The control itself operates with the shipped Colang 2.x rails. The fix lands with the inference-backend kernel change, before CAGE_INFERENCE_BACKEND=none exists.
    • POAM-2026-120 gains an instance in the consensus path. Finance's HIGH_VALUE_TRADE high-stakes override is asserted present by a test but has never fired, so consensus is triggered by amount alone.
  • filelock 4.2.4, langgraph-api 0.16.9 and langgraph-cli 0.4.40 clear four dependency advisories (pyproject.toml, uv.lock).

  • authlib 1.9.1 and msgpack 1.2.4 clear eight dependency advisories (uv.lock only). They were published after main's last clean dependency scan, and from then on the scan failed on every pull request.

    Nothing in src/ imports either package, and no other locked package changed.

Full Changelog: v4.0.0...v4.1.0