Module ethics ai Architecture

Architecture - Ethics AI Module

Overview

The ethics_ai module composes philosophy profile ingestion, argument persistence, discourse orchestration, context assembly, evaluation, and plugin lifecycle wiring into a coherent ethics reasoning runtime for ThemisDB.

The current production mode is SELECTION_ONLY (Top-N school pre-selection via EthicsSelectionRouter). The Layered Discourse Model (LDM) adds LAYERED_FULL and LAYERED_FAST modes that satisfy Habermas participatory fairness while remaining computationally tractable. See FUTURE_ENHANCEMENTS.md and docs/research/ethics_discourse_process_equality.md for design details.

Main Execution Planes

1. Profile and Configuration Plane

• profile loading, validation, and registry/index operations
• philosophy selection and routing across school sets
• LDM extension: DiscourseMode selection; DiscourseOrchestratorPlan generation assigning all N schools to Ebene-1 with equal initial weight w₀ = 1/N

2. Discourse and Synthesis Plane

• debate initialization/continuation and decision generation
• cross-school tension resolution and synthesis support
• LDM extension (Ebene-1): parallel equal-weight initial scoring for all N schools; verdict ∈ {PROHIBIT, PERMIT, CONDITIONAL, ABSTAIN} per school
• LDM extension (Ebene-2): taxonomy-class cluster discourse (K≤6 clusters) along four structural tension axes; O(K²·R) instead of O(N²·R)
• LDM extension (Ebene-3): convergence-counting MetaVerdict with cross_cultural_flag, minority_dissent[], and positivrechtlichem Legal-DB-Grounding

3. Memory/Context Plane

• argument and decision storage/retrieval contracts
• RAG context assembly, prior-round compression, and matrix synthesis
• LDM extension: ClusterPosition persistence per Ebene-2 cluster; EpisodicMemoryEntry for inter-cluster tension outcomes

4. Evaluation and Operations Plane

• ethics scoring and metrics emission
• plugin lifecycle integration and runtime observability behavior
• LDM extension: MetaVerdict audit fields (participating_schools, convergence_score, discourse_mode) for EU AI Act Art. 13 transparency compliance

5. Cultural Self-Reflection Plane (LDM Mirror-School-Modus)

• non-western schools (islamische_ethik, konfuzianismus, buddhistische_ethik, juedische_bioethik) participate as structural self-reflection mirrors
• activated per-domain via cross_cultural_sensitivity policy
• output: minority_dissent[] in MetaVerdict — always present in audit trail
• cost: ≤ 1 LLM inference step per mirror school (parallel to Ebene-2)

Core Contracts

Contract	Behavior
profile contract	explicit profile validation/load and lookup behavior
discourse contract	deterministic debate and decision orchestration semantics
process equality contract (LDM)	all N schools participate in Ebene-1 with equal initial weight
audit completeness contract (LDM)	MetaVerdict lists all N schools incl. ABSTAIN
context contract	bounded argument memory and context retrieval behavior
evaluation/ops contract	explicit scoring, metrics, and plugin lifecycle surfaces
legal grounding contract	MetaVerdict grounding uses DB citations, never LLM paraphrase

LDM Discourse Mode Overview

DiscourseMode::SELECTION_ONLY   (current production)
  └─ EthicsSelectionRouter: Top-N pre-selection
  └─ Discourse: top_n schools only
  └─ MetaVerdict: weighted aggregate

DiscourseMode::LAYERED_FULL     (implemented)
  └─ Ebene-1: all N schools, parallel, equal w₀=1/N         O(N)
  └─ Ebene-2: K≤6 clusters × 3 rounds, tension-axis routing  O(K²·R)
  └─ Ebene-3: convergence-count + legal-DB grounding          O(1)+lookup
  └─ Mirror: non-western schools, 1 step parallel to Ebene-2
  └─ MetaVerdict: convergence_score + cross_cultural_flag + minority_dissent

DiscourseMode::LAYERED_FAST     (implemented)
  └─ Ebene-1: all N schools parallel (identical to LAYERED_FULL)
  └─ Ebene-2: axis-1 only (Kant↔Utilitarismus) + domain-relevant cluster
  └─ Ebene-3: MetaVerdict from Ebene-1 convergence + legal-DB
  └─ Target: P95 ≤ 1.2 s end-to-end

Failure Semantics

• invalid profile or school selection paths fail with structured status/error states.
• missing storage/context inputs produce explicit non-silent failure behavior.
• lifecycle misuse (uninitialized plugin paths) fails deterministically.
• LDM: LLM timeout in Ebene-1 → school receives ABSTAIN (fail-closed); never silent drop.
• LDM: Legal-DB unavailable → MetaVerdict emitted without grounding with explicit observable flag.
• LDM: All N schools ABSTAIN → DISSENT MetaVerdict; never silently promoted to SELECTION_ONLY.

Sourcecode Verification (Module: ethics_ai/architecture)

• Verified files:
  • src/ethics_ai/ethics_ai_plugin.cpp
  • src/ethics_ai/discourse_engine.cpp
  • src/ethics_ai/argument_store.cpp
  • src/ethics_ai/philosophy_loader.cpp
  • src/ethics_ai/rag_context_engine.cpp
  • src/ethics_ai/ethics_evaluator.cpp
  • src/ethics_ai/ethics_profile_registry.cpp
  • src/ethics_ai/ethics_selection_router.cpp
  • src/ethics_ai/discourse_orchestrator.cpp (DiscourseMode + DiscourseOrchestratorPlan)
  • src/ethics_ai/cluster_discourse_engine.cpp (Ebene-2 cluster routing)
  • src/ethics_ai/meta_verdict_builder.cpp (Ebene-3 convergence-count + legal grounding)
  • src/ethics_ai/mirror_school_handler.cpp (Mirror-School-Modus)
• Verified architecture claims:
  • explicit profile/discourse/context/evaluation planes
  • bounded deterministic failure behavior in lifecycle and profile paths
  • module-local orchestration for ethics reasoning runtime flows
  • LDM Ebene-1/2/3 and Mirror-School-Modus implemented and verified

---

ThemisDB 1.9.0-beta · Home · Module-Index · GitHub · Issues