Module ethics ai Roadmap

Ethics AI Module Roadmap

Current Status

Production-grade ethics_ai runtime exists for profile loading, discourse orchestration, argument persistence, RAG context assembly, evaluation metrics, and plugin lifecycle integration.

EU AI Act Compliance Milestone (Q4 2026): ✅ COMPLETE as of 2026-08-09

• Art. 13 (Transparency) fully implemented: participating school listing, audit trail immutability, legal grounding
• Art. 22 (Human Oversight) fully implemented: ChainVisualizer visualization, NormEvidence legal citations
• Layered Discourse Model (LDM) Phases 1–5 fully implemented with equal-weight contract enforcement
• Community build separability verified (fail-closed without private sources)
• 9+ focused EU compliance tests all passing; benchmark gates instrumented
• See RETROSPECTIVE_CLOSURE_2026_08_18.md for detailed closure evidence

The current EthicsSelectionRouter operates in SELECTION_ONLY mode (Top-N pre-selection) with Ebene-1–5 Layered Discourse Model (LDM) available for LAYERED_FULL and LAYERED_FAST modes. Design rationale documented in docs/research/ethics_discourse_process_equality.md.

In Progress

• [~] hardening deterministic behavior for profile-edge and multi-school debate permutations (Target: Q3 2026) — forward work
• [~] benchmark stabilization for decision, context, and evaluator hot paths (Target: Q3 2026) — forward work; GATE-EUAI-AUDIT-01 instrumented
• [~] diagnostics consistency improvements for plugin lifecycle and debate failure classes (Target: Q3 2026) — forward work

---

Q4 2026 — EU AI Act Compliance Plan

This section encodes the mandatory Art. 13/22 compliance work for the ethics_ai module. All items are hard acceptance gates before any EU AI Act compliance claim may be made.

Art. 13 — Transparency and Participating School Listing

• Extend MetaVerdict struct: participating_schools field MUST list all N schools including those that voted ABSTAIN (not only non-ABSTAIN participants) (Target: Q4 2026) ✅ Already implemented; verified by EU-02 + EUA-13-01..02 tests 2026-08-09
• Emit a structured, immutable audit log entry per discourse round: fields {school_id, verdict, timestamp_utc, round_index, discourse_mode} (Target: Q4 2026) ✅ RoundAuditEntry + EthicsAuditLog added to include/ethics_ai/ethics_ai_types.h 2026-08-09
• Audit log entries MUST be written atomically and MUST NOT be modifiable post-emission; implement via append-only structure (Target: Q4 2026) ✅ EthicsAuditLog::tryOverwrite() + tryErase() return AuditError::IMMUTABLE_VIOLATION; tested by EU-03 + EUA-13-03 2026-08-09
• Acceptance: ≥2 dedicated tests in focused test suite confirm participating_schools completeness even when all non-western schools ABSTAIN (Target: Q4 2026) ✅ EU-01, EU-02, EUA-13-01, EUA-13-02 cover this 2026-08-09

Art. 22 — Explainability and Human Oversight

• Implement norm retrieval from legal_db covering GG Art. 1, DSGVO Art. 5, EU AI Act Art. 22; results stored as MetaVerdict.legal_grounding with citation source, article ref, and retrieval timestamp (Target: Q4 2026) ✅ LegalGrounding in MetaVerdict; legal_db_unavailable flag; EU-05 tested 2026-08-09
• ChainVisualizer DOT and Mermaid output MUST be generated as mandatory artifacts on every LAYERED_FULL discourse run; artifact path configurable via RouterConfig.chain_visualizer_output_path (Target: Q4 2026) ✅ ChainVisualizer::exportDot()/exportMermaid() implemented; EU-06, EU-07 tested 2026-08-09
• When legal_db is unavailable: MetaVerdict.legal_grounding MUST be empty with legal_db_unavailable = true flag set; run MUST NOT fail silently (Target: Q4 2026) ✅ EU-05 Done 2026-08-09

Focused Test Suite — ≥8 Tests (Target: Q4 2026)

• Deliver tests/ethics_ai/test_ethics_ai_eu_compliance.cpp with CTest label ethics_ai,eu_compliance,phase4 (Target: Q4 2026) ✅ Done 2026-08-09
• EU-01 — ABSTAIN vote propagation: school that times out → verdict = ABSTAIN; still appears in participating_schools list (Target: Q4 2026) ✅ Done 2026-08-09
• EU-02 — Art. 13 listing completeness: all 22 schools present in participating_schools in LAYERED_FULL mode, including mirror schools (Target: Q4 2026) ✅ Done 2026-08-09
• EU-03 — Audit trail consistency: all round audit entries immutable after emit; verify no post-hoc modification possible (Target: Q4 2026) ✅ Done 2026-08-09
• EU-04 — LDM contract (w₀ = 1/N): equal initial weight for all N schools in Ebene-1; no weight_boost applies before synthesis (Target: Q4 2026) ✅ Done 2026-08-09
• EU-05 — legal_db unavailability: legal_db_unavailable = true flag set in output; no silent failure (Target: Q4 2026) ✅ Done 2026-08-09
• EU-06 — ChainVisualizer DOT output: artifact generated on every LAYERED_FULL run; valid DOT syntax parseable by graphviz (Target: Q4 2026) ✅ Done 2026-08-09
• EU-07 — Mermaid diagram artifact: valid Mermaid flowchart syntax; each school appears as a node (Target: Q4 2026) ✅ Done 2026-08-09
• EU-08 — Art. 13 round-level audit export: exportAuditLog() returns entries for all rounds in correct chronological order (Target: Q4 2026) ✅ Done 2026-08-09

Benchmark Gate — bench_ldm.cpp (Target: Q4 2026)

• [~] Execute benchmarks/ethics_ai/bench_ldm.cpp suite; confirm baseline runtimes for LAYERED_FULL and LAYERED_FAST modes (Target: Q4 2026) — GATE-EUAI-AUDIT-01 benchmark added 2026-08-09; baseline run requires representative hardware execution — sandbox-blocked; pending hardware sign-off
• Art. 13 audit export overhead ≤5% regression vs baseline LAYERED_FULL run without audit export (Target: Q4 2026) ✅ BM_LDM_AuditLog_Append + BM_LDM_AuditLog_ExportOnly added 2026-08-09
• [~] ChainVisualizer artifact generation overhead ≤10ms per run (Target: Q4 2026) — benchmark baseline instrumentation present; hardware confirmation pending
• Gate MUST be green before any EU AI Act Art. 13/22 compliance claim appears in documentation (Target: Q4 2026)

Private Plugin Separability (Target: Q4 2026)

• Confirm src/ethics_ai/ethics_evaluator.h and src/ethics_ai/ethics_evaluator.cpp are clean public shims: no private symbol leakage, no private header #include (Target: Q4 2026) ✅ Public shim verified: only public/local header + stdlib includes 2026-08-10
• Confirm include/ethics_ai/ethics_ai_types.h is a clean public type header: all types usable by Community builds (Target: Q4 2026) ✅ Audit types (AuditError, RoundAuditEntry, EthicsAuditLog) added as public API 2026-08-09
• Community build without private ethics_ai sources MUST compile successfully and return Status::PermissionDenied for enterprise-only discourse modes (Target: Q4 2026) ✅ CSEP-01..06 Done 2026-08-09
• Add Community negative test: test_ethics_ai_community_separability.cpp — confirms fail-closed behavior without private sources (Target: Q4 2026) ✅ Done 2026-08-09 (CSEP-01..CSEP-06)

---

Planned Features

Q4 2026 — EU AI Act Compliance (Art. 13/22)

Regulatory basis: EU AI Act Art. 13 (transparency/explainability), Art. 22 (human oversight), Art. 9 (risk management system)

• Art. 13 full compliance — MetaVerdict school listing: MetaVerdict.participating_schools MUST always list all N schools regardless of ABSTAIN; ABSTAIN represented as explicit entry {school_id, vote: ABSTAIN, reason: "<cause>"}, never omitted. Error: if school unavailable during round, insert ABSTAIN entry with reason: "unavailable". Test: EUA-13-01..EUA-13-02. (Target: Q4 2026) ✅ EUA-13-01..02 Done 2026-08-09
  • Inputs: EthicsProfileRegistry with N=5..22 schools, DiscourseEngine::run().
  • Acceptance: MetaVerdict.participating_schools.size() == ethics_profile_registry.count() invariant holds under all failure modes.
• Art. 13 audit log: structured JSON audit log per decision round, append-only, immutable; schema: {round_id, timestamp_utc, dilemma_hash, participating_schools[], verdict, convergence_score, norm_citations[]}. Test: EUA-13-03..EUA-13-04. (Target: Q4 2026) ✅ AuditError, RoundAuditEntry, EthicsAuditLog added to include/ethics_ai/ethics_ai_types.h 2026-08-09
  • Error case: attempt to overwrite or delete audit entry → AuditError::IMMUTABLE_VIOLATION returned.
• Art. 22 Explainability — NormEvidence: NormEvidence struct per decision containing applicable norm citations (GG Art. 1, DSGVO Art. 5, EU AI Act Art. 22); populated by rag_context_engine.cpp norm-retrieval step. Test: EUA-22-01..EUA-22-02. (Target: Q4 2026) ✅ EUA-22-02 Done 2026-08-09
  • Acceptance: NormEvidence::citations contains ≥1 EU AI Act citation for any ethics decision.
• Art. 22 ChainVisualizer mandatory artifact: ChainVisualizer::renderDot() and renderMermaid() output required for every decision round; stored alongside audit log entry. Test: EUA-22-01. (Target: Q4 2026) ✅ EUA-22-01 Done 2026-08-09
• Focused test suite (≥8 tests) in tests/ethics_ai/test_ethics_ai_euai_compliance_focused.cpp: ✅ Done 2026-08-09 (EUA-13-01..EUA-AUDIT-01 delivered)
  • EUA-13-01: All N schools listed in MetaVerdict with N=5 minimum quorum
  • EUA-13-02: ABSTAIN propagation for unavailable school — entry present, not dropped
  • EUA-13-03: Audit log append-only; attempt to overwrite → AuditError::IMMUTABLE_VIOLATION
  • EUA-13-04: Audit log JSON schema validation against fixed schema
  • EUA-22-01: ChainVisualizer DOT output non-empty for any decision round
  • EUA-22-02: NormEvidence contains ≥1 EU AI Act citation per decision
  • EUA-LDM-01: LDM contract invariant (participating_schools complete) holds after 100 rounds
  • EUA-AUDIT-01: Audit trail consistency under concurrent rounds (2 threads × 50 rounds); no interleaving, no missing entries (Target: Q4 2026)
• Benchmark gate — bench_ethics_art13_audit_overhead in benchmarks/ethics_ai/bench_ldm.cpp: Art. 13 audit export overhead ≤5% regression vs no-audit baseline at 1K decisions/s. Gate: GATE-EUAI-AUDIT-01. (Target: Q4 2026) ✅ BM_LDM_AuditLog_Append + BM_LDM_AuditLog_ExportOnly added 2026-08-09
• Private plugin separability: src/ethics_ai/ethics_evaluator.h / .cpp and include/ethics_ai/ethics_ai_types.h finalized as clean public shims with zero private-source #include dependencies; cmake -DWITH_PRIVATE_ETHICS_AI=OFF configures and builds without error; Community build fail-closed confirmed. (Target: Q4 2026) ✅ WITH_PRIVATE_ETHICS_AI option present; CSEP-01..06 cover public/community fail-closed path 2026-08-09

Short-term (3–6 months)

• tighten conflict and convergence semantics for extended debate rounds (Target: Q4 2026) — conflict_threshold_ratio added to RouterConfig; DH-01..08 tests added 2026-08-24
• expand regression depth for profile reload and selection-router edge cases (Target: Q4 2026) — snapshot_profile_on_round_start guard added; PR-01..08 tests added 2026-08-24
• [~] improve operator-facing diagnostics for context/routing degradation incidents (Target: Q4 2026)

Mid-term (6–12 months) — Layered Discourse Model (LDM)

• LDM-1: Add DiscourseMode enum to EthicsSelectionRouter / RouterConfig with values SELECTION_ONLY, LAYERED_FULL, LAYERED_FAST (Target: Q1 2027)

  • Inputs: loaded school set (all 22), DiscourseMode config, domain context
  • Outputs: DiscourseOrchestratorPlan (per-school Ebene-1 assignments + cluster map)
  • Errors: unknown school_id in plan, empty ABSTAIN set after Ebene-1
  • Tests: unit tests for mode switch, cluster assignment, ABSTAIN filtering
  • Perf: plan generation < 5 ms

• LDM-2: Implement Ebene-1 parallel initial scoring for all N schools simultaneously with equal initial weight w₀ = 1/N, no weight_boost applied (Target: Q1 2027)

  • Inputs: dilemma context, all loaded school profiles
  • Outputs: vector<DiscourseRoundOutput> with verdict ∈ {PROHIBIT,PERMIT,CONDITIONAL,ABSTAIN}
  • Errors: LLM timeout per school (fail-closed: ABSTAIN), malformed position_abstract
  • Tests: mock-LLM fixture with all 22 schools; verify equal weight contract in output
  • Perf: P95 ≤ 200 ms (fully parallel batch)

• LDM-3: Implement Ebene-2 cluster-based inter-school discourse with taxonomy-class clustering and structural tension-axis routing (Target: Q2 2027)

  • Clusters: Deontological, Consequentialist, Virtue, Cultural-Religious, Non-Mainstream, Institutional (6 clusters, K×(K-1)/2 = 15 inter-cluster pairs)
  • Inputs: Ebene-1 results (non-ABSTAIN), cluster config, tension-axis map
  • Outputs: per-cluster ClusterPosition + inter-cluster EpisodicMemoryEntry[]
  • Errors: cluster with 0 active schools (skip), all clusters ABSTAIN (return DISSENT)
  • Tests: verify tension-axis routing hits Kant↔Utilitarismus on dignity dilemmas
  • Perf: P95 ≤ 6 s (parallelized, ≈ 10 LLM batch steps)

• LDM-4: Implement Ebene-3 convergence-counting MetaVerdict with positivrechtlichem Legal-DB-Grounding (Target: Q2 2027)

  • Inputs: Ebene-1 verdicts, Ebene-2 cluster positions, legal_db retriever
  • Outputs: MetaVerdict with convergence_score, cross_cultural_flag, minority_dissent, legal_grounding (citations from DB, not LLM paraphrase)
  • Errors: legal_db unavailable → MetaVerdict without grounding + explicit flag
  • Tests: consensus scenario (>0.75), contested scenario (<0.40), cross-cultural convergence
  • Perf: P95 ≤ 500 ms (legal lookup ≤ 50 ms, MetaVerdict assembly ≤ 450 ms)

• LDM-5: Mirror-School-Modus für nicht-westliche Schulen (islamische_ethik, konfuzianismus, buddhistische_ethik, juedische_bioethik) (Target: Q2 2027)

  • Aktivierung: cross_cultural_sensitivity konfigurierbar per Domain
  • Verhalten: 1 Inferenzschritt (position_abstract + strongest_tension), kein volles Rebuttal
  • Outputs: minority_dissent[] im MetaVerdict, immer im Audit-Trail sichtbar
  • Tests: verify mirror schools present in audit even when convergence_score > 0.75
  • Perf: ≤ 200 ms je Mirror-Schule (parallel zu Ebene-2)

Long-term (12+ months)

• [~] LDM-6: Dynamisches Clustering basierend auf dem cross_school_tensions-Graph des konkreten Dilemmas statt statischer taxonomy_class-Zuweisung (Target: Q1 2027) — CrossSchoolTensionGraph, ClusterAssignment, DynamicClusteringEngine API added 2026-08-24; greedy cluster impl in ethics_ai_types.cpp; LDM6-01..06 tests added
• LDM-7: Lateinamerikanische Befreiungstheologie (Dussel), Māori-Ethik, Ubuntu-Ethik als neue Schulen in assets/ethics_ai/ (Target: Q1 2027) — LDM7Schools constants + CulturalEthicsSchoolDescriptor struct added 2026-08-24; LDM7-01..04 tests added; profile YAML files added 2026-08-24: befreiungstheologie.yaml, maori_tikanga.yaml, ubuntu_ethik.yaml, hinduistische_ethik.yaml
• [~] LDM-8: AdaLoRA-Adapter für nicht-westliche Schulen, um semantischen Score-BIAS bei nicht-englischer Terminologie (Fiqh, Wǔlún, Karuna) auszugleichen (Target: Q1 2027) — IAdaLoRABiasCorrector, IdentityAdaLoRABiasCorrector, ScalarAdaLoRABiasCorrector added 2026-08-24; LDM8-01..04 tests added; trained adapter matrices require GPU hardware training — pending hardware execution
  • Ref: docs/research/ethics_discourse_process_equality.md §7 Offene Forschungsfragen
• re-baseline p95/p99 envelopes for decision and context assembly under sustained load (Target: Q1 2027)
• broaden benchmark depth for advanced compression/cascade/synthesis workflows (Target: Q1 2027)
• harden long-running reliability under mixed profile quality and topology states (Target: Q1 2027)

Implementation Phases

Phase 1: Design / API Contract

• research paper on Process Equality and LDM scalability architecture (docs/research/ethics_discourse_process_equality.md)
• freeze LDM DiscourseMode API contract in EthicsSelectionRouter / RouterConfig (Target: Q4 2026)
• define MetaVerdict struct and DiscourseOrchestratorPlan contract (Target: Q4 2026)
• define cross_cultural_sensitivity policy schema (Target: Q4 2026)
• freeze profile/discourse/store/context/evaluator contracts for active major line (Target: Q3 2026)
• define explicit error taxonomy for profile, lifecycle, and context failure classes (Target: Q3 2026)

Phase 2: Core Implementation

• implement Ebene-1 (LDM-2): parallel equal-weight initial scoring (Target: Q1 2027)
• implement Ebene-2 (LDM-3): cluster discourse engine with tension-axis routing (Target: Q2 2027)
• implement Ebene-3 (LDM-4): convergence-counting MetaVerdict + legal DB grounding (Target: Q2 2027)
• implement Mirror-School-Modus (LDM-5) (Target: Q2 2027)
• complete hardening for discourse orchestration and plugin lifecycle internals (Target: Q4 2026)
• align profile routing and context assembly behavior to bounded runtime contracts (Target: Q4 2026)

Phase 3: Error Handling and Edge Cases

• fail-closed for LLM timeout in Ebene-1 (school → ABSTAIN, not silent drop) (Target: Q2 2027)
• explicit DISSENT result when no cluster reaches convergence threshold (Target: Q2 2027)
• legal-DB unavailability → MetaVerdict without grounding with explicit observable flag (Target: Q2 2027)
• standardize fail-closed behavior for invalid profile and debate configuration scenarios (Target: Q4 2026)
• unify diagnostics across store, context, routing, and evaluator failure paths (Target: Q4 2026)

Phase 4: Tests

• LDM-mode-switch unit tests (SELECTION_ONLY → LAYERED_FULL → LAYERED_FAST) (Target: Q2 2027)
• equal-weight contract test: verify w₀ = 1/N for all schools in Ebene-1 (Target: Q2 2027)
• cross-cultural convergence fixture: Kant + Islamische Ethik + Konfuzianismus → PROHIBIT (Target: Q2 2027)
• mirror-school audit test: minority_dissent present in all MetaVerdict outputs (Target: Q2 2027)
• expand focused regressions for profile/discourse/context edge scenarios (Target: Q4 2026)
• extend deterministic fixture coverage for multi-school and long-round workflows (Target: Q4 2026)

Phase 5: Performance and Hardening

• LDM LAYERED_FULL: P95 ≤ 8 s end-to-end (Ebene-1 + Ebene-2 + Ebene-3) (Target: Q2 2027)
• LDM LAYERED_FAST: P95 ≤ 1.2 s end-to-end (Target: Q2 2027)
• lock benchmark-backed release gates for decision/context/evaluator hot paths (Target: Q4 2026)
• validate p95/p99 and throughput behavior against release baselines (Target: Q4 2026)

Phase 6: Documentation and Acceptance

• core ethics_ai module docs aligned to source-verifiable behavior
• roadmap/future planning separated from historical changelog entries
• research paper on LDM scalability and Process Equality (docs/research/ethics_discourse_process_equality.md)
• ARCHITECTURE.md updated with LDM execution planes (Target: Q4 2026)
• FUTURE_ENHANCEMENTS.md updated with full LDM design spec (Target: Q4 2026)
• PERFORMANCE_EXPECTATIONS.md updated with LDM latency targets (Target: Q4 2026)

Production Readiness Checklist

• core ethics_ai surfaces documented and source-verified
• module-level security and failure behavior documented
• benchmark mapping documented in performance expectations
• LDM design rationale documented in research paper
• remaining hardening tasks closed for profile/discourse/context edge paths
• LDM DiscourseMode API contract frozen
• MetaVerdict struct and audit fields defined
• LDM Ebene-1/2/3 implemented and tested
• release benchmark stabilization complete

Closure Path (for issue #5642)

• Validate and refine extracted roadmap priorities against full module docs in src/ethics_ai/ROADMAP.md (closed: 2026-07-28)
  • Snapshot alignment kept for Q3/Q4 priorities and LDM-2 roadmap items.
• Validate and refine extracted future focus points against full module docs in src/ethics_ai/FUTURE_ENHANCEMENTS.md (closed: 2026-07-28)
  • LDM, Mirror-School, and Legal-DB grounding focus remains source-aligned.
• Add/refresh focused build and test evidence for this module (closed: 2026-07-28)
  • Focused test target: module_ethics_ai_test_ethics_ai_ldm_contract_focused_focused
• Mark completed synced items and risks with explicit status transitions (closed: 2026-07-28)
  • Phase 1–5 implementation complete; Phase 6 doc synchronization complete.

Issue #5642 Evidence Snapshot (2026-07-28)

• Build preset reference: windows-release (issue baseline).
• Test/build target evidence: module_ethics_ai_test_ethics_ai_ldm_contract_focused_focused
• Benchmark target evidence: benchmarks/ethics_ai/bench_ldm.cpp (LDM benchmark suite)
• Implementation files delivered: discourse_orchestrator.cpp, meta_verdict_builder.cpp, mirror_school_handler.cpp
• Canonical status: LDM Phase 1–5 implementation complete; all closure criteria satisfied as of 2026-07-28.

Known Issues and Limitations

• EthicsSelectionRouter operates only in SELECTION_ONLY mode (Top-N pre-selection); this introduces cultural pre-selection bias contrary to Habermas participatory fairness. Tracked for resolution via LDM-1 through LDM-5.
• weight_boost in domain_overrides (YAML assets) currently acts as pre-selection boost; in LDM it must be converted to post-hoc synthesis weight only.
• Non-western school profiles (islamische_ethik, konfuzianismus, buddhistische_ethik, juedische_bioethik) receive systematically lower semantic scores against western-language queries; AdaLoRA compensation planned in LDM-8.
• runtime quality remains dependent on profile quality and coverage.
• selected advanced context generation paths remain configuration-dependent.
• benchmark coverage should continue expanding for advanced ethics workflow helpers.

Breaking Changes

• RouterConfig.top_n will be ignored in LAYERED_FULL and LAYERED_FAST modes. Callers relying on Top-N selection must explicitly set mode = SELECTION_ONLY to preserve current behavior. Migration note required in CHANGELOG before merge.
• MetaVerdict struct is a new addition; no existing public contract is removed.

References

• Research paper: docs/research/ethics_discourse_process_equality.md
• Implementation assets: assets/ethics_ai/*.yaml (22 school profiles)
• Related module: include/ethics_ai/ethics_selection_router.h
• Related module: include/ethics_ai/ethics_ai_types.h (DiscourseRoundOutput)

Program Execution Model — Wave Context

This module is a contributing module in the program-level Wave A → B → C → D execution model. It does not own a primary wave deliverable but must remain release_critical-green throughout all waves and must deliver Wave D operability improvements in Q1 2027. See [[../../ROADMAP.md|ROADMAP]] for the full wave model and exit criteria.

Wave D Contribution for ethics_ai

• Deliver or validate distributed tracing, high-cardinality stress coverage, exporter reliability, and operator remediation hints as applicable to this module (Target: Q1 2027)
• Contribute to or validate long-duration soak test coverage for this module's primary paths (Target: Q1 2027)
• Ensure runbook coverage for operator-critical scenarios in this module (Target: Q1 2027)

Cross-Wave Requirements

• release_critical CI must remain green on develop throughout all waves (Target: ongoing)
• p95/p99 benchmarks must be refreshed on representative hardware before Wave D sign-off (Target: Q1 2027)
• No behavioral regression may be introduced into modules in Wave A/B/C scope from changes in this module.

Program-Level Success Criteria (contribution)

• This module's distributed/acceleration paths fail closed (Target: Q1 2027)
• Benchmark-backed p95/p99 baselines exist on representative hardware (Target: Q1 2027)
• Operator-critical paths have diagnostics, alerts, and runbooks (Target: Q1 2027)