Eve library

Oshun Domains — Separation of Concerns & Duplication Audit

Guide

Revision 1

On this page

Date: 2026-04-06 Scope: All 42 domains under DOMAINS/ Method: Cross-referencing features.md, architecture.md, and specifications.md for every domain; verifying declared cross-domain dependencies against actual library imports and architecture diagrams


Table of Contents#

  1. Executive Summary
  2. Methodology
  3. Domains With Clean Separation
  4. Overlap Cluster Analysis
  5. Cross-Domain Dependency Health
  6. Root Cause Analysis
  7. Recommendations
  8. Appendix A — Full Domain Purpose Index
  9. Appendix B — Duplication Severity Matrix

1. Executive Summary#

Across 42 domains, domain purposes are well-separated — each domain has a clear, distinct mission with no ambiguity about what belongs where. However, 8 functional capabilities are duplicated across multiple domains instead of being shared through a canonical owner.

At a Glance#

Severity Count Clusters
Critical Duplication 2 IRT/Psychometrics (Metis + Mnemosyne), Persona Modeling (Psyche + Calliope + Hathor)
Significant Duplication 2 Voice/TTS (Psyche + Lilith vs. Euterpe), Blockchain/Web3 (Lilith vs. Aje)
Moderate Duplication 1 RAG/Knowledge (Lilith + Veritas vs. Sophia)
Minor / Justified 1 Meditation (domain-specific needs)
Architecture Doc Issue 1 Galatea claims Aja consumption but declares no dependency
Clean (No Action) 1 Content Generation (properly delegated via Isis)

Key Findings#

  • 28 of 42 domains have completely unique, non-overlapping missions with zero duplication concerns
  • Lilith is the primary source of duplication — its 80+ microservice architecture builds independent TTS, RAG, blockchain, and meditation stacks rather than consuming canonical domains
  • Content generation (Isis → Yemaya, Calliope) is the model of how cross-domain consumption should work
  • Psychometrics (Metis + Mnemosyne) is the most clear-cut case of unnecessary duplication: identical IRT/CAT/ItemBank implementations with zero shared code

2. Methodology#

Process#

  1. Read the opening sections and feature inventories of all 42 features.md files to map each domain's purpose and key capabilities
  2. Identified 8 functional capability clusters where 2+ domains describe overlapping functionality
  3. For each cluster, read the full architecture.md of every involved domain, focusing on:
    • Library inventories and what each library does
    • Cross-domain dependency declarations
    • Actual import paths and adapter patterns
    • Technology choices (vector DBs, TTS providers, blockchain libraries)
  4. Compared declared dependencies against actual architecture to identify gaps

What Constitutes Duplication#

Classification Definition
Critical Identical algorithms/data models implemented independently in 2+ domains with zero sharing. Creates divergence risk and double maintenance.
Significant Same capability category built independently when a canonical domain exists and is consumed by other domains — but not by the duplicating domain.
Moderate Same capability category with different technology choices that may be justified by domain-specific requirements, but where shared abstractions could still reduce redundancy.
Minor / Justified Overlapping functionality where domain-specific context (yoga-specific breathing vs. generic meditation) makes independent implementation reasonable.
Clean Multiple domains use the same capability but properly delegate to the canonical owner.

3. Domains With Clean Separation#

These 28 domains have unique, non-overlapping missions with no duplication concerns:

Infrastructure & Platform#

Domain Purpose
Shared 30 cross-cutting infrastructure libraries: auth, database, caching, events, logging, metrics, tracing, HTTP client, storage, security, AI integration
Contracts Canonical entity schemas, Zod validation, event envelopes for cross-domain communication
OpenAPI Centralized OpenAPI 3.1 specification management, type generation, breaking change detection
Proto Protocol Buffer definitions for gRPC inter-service communication
Oshun Platform shell, cross-domain orchestration, voice-first assistant, domain adapters
Neith Sovereign Rust runtime kernel: async execution, custom allocators, cryptography, hardware abstraction

Creative & Production#

Domain Purpose
Bellona Build bridges connecting creative domains to game engines (Blender, Godot, Unreal, Unity)
Maya Infinite virtual universe / metaverse game engine: Lumen GI, Nanite geometry, multi-physics, MMO
Uzume Live entertainment production: lighting, audio, video, rigging, DMX/Art-Net/sACN/OSC/MIDI protocols
Aphrodite Adult entertainment streaming: RTMP ingest, haptic devices, token economy, performer safety

Knowledge & Science#

Domain Purpose
Nyx Cosmic observatory: star catalogs (Gaia DR3 with 1.8B sources), orbital mechanics, telescope control
Nisaba Ancient text analysis: 18 writing systems, morphological analysis, textual criticism, paleography
Airmid Evidence-based phytotherapy: computational phytochemistry, ADMET prediction, network pharmacology
Sophia Knowledge management: document ingestion, semantic search, RAG pipeline, knowledge graph, citations

Fashion, Beauty & Retail#

Domain Purpose
Aglaea AI fashion/beauty platform: 150+ skin biomarkers, body type classification, virtual try-on, trend forecasting
Galatea Humanoid robotics for fashion retail: ISO 13482 safety, kinematics, choreographed shows, fleet management

Agriculture, Food & Home#

Domain Purpose
Asase Ghana commercial agriculture: agro-ecological zones, cold chain logistics, export operations, COCOBOD compliance
Demeter Home gardening: companion planting, IoT sensor networks, hydroponics, aquaponics, AI plant identification
Hestia Culinary intelligence: recipe science, guided cooking, smart kitchen IoT, nutritional analysis, sustainability
Seshat Dwelling arts & craftsmanship: woodworking, CNC/laser/3D fabrication, interior design, smart home IoT

Business, Governance & News#

Domain Purpose
Maat Autonomous organization OS for West African business portfolios: strategic intelligence, financial modeling
Veritas Autonomous AI news agency for Ghana/West Africa: 300+ sources, NLP analysis, fact-checking, multimedia
Kuanyin Platform-wide ethics, safety, and moderation: 50+ harm categories, rehabilitation, merit/karma systems

Learning & Personal Development#

Domain Purpose
Arete Personal development: habit formation, goal setting (OKR), journaling (CBT), time management, gamification
Meditation Platform-agnostic meditation engine: 6 shared libraries (player, timer, session, breathing, progress, offline)
Tara Meditation app (PWA + native): consumes all 6 @oshun/meditation-* libraries, adds platform layers

Other#

Domain Purpose
Oya Drone / autonomous aerial systems: 3D math, flight states, mission planning, PX4/ArduPilot integration
Aje Canonical blockchain/DeFi/Web3 infrastructure: wallets, multi-chain, DeFi, NFTs, smart contracts, ZK proofs

4. Overlap Cluster Analysis#

4.1 Assessment / Psychometrics (IRT, CAT)#

Severity: CRITICAL Domains: Metis, Mnemosyne

What Each Domain Implements#

Capability Metis (@metis/assessment) Mnemosyne (@mnemosyne/core)
IRT 1-Parameter (Rasch) irt1PL irt1PL
IRT 2-Parameter irt2PL irt2PL
IRT 3-Parameter irt3PL irt3PL
Computerized Adaptive Testing CAT with max-information selection CAT with max-information selection
Item Bank ItemBank class with psychometric stats ItemBank class with psychometric stats
Ability Estimation EAP (Expected A Posteriori) MLE + EAP
Fisher Information Yes Yes
Portfolio Assessment Yes PortfolioTracker
Peer Assessment Yes PeerAssessmentManager
Certification N/A CertificationManager
Spaced Repetition No SM-2, FSRS, HLR, Ebbinghaus, LECTOR

Cross-Domain Dependencies#

text
Metis ──── (no dependency) ──── Mnemosyne

Both implement IRT/CAT/ItemBank independently.
Zero shared library. Zero cross-domain imports.

Why This Is Critical#

IRT is mathematically precise. Two independent implementations of the same psychometric algorithms creates:

  • Numerical divergence risk — different floating-point handling, convergence criteria, or edge-case behavior could produce different ability estimates for the same response pattern
  • Double maintenance — when psychometric best practices evolve (e.g., new estimation methods), both implementations must be updated independently
  • Item bank incompatibility — items calibrated in Metis cannot be used in Mnemosyne and vice versa, because the IRT parameters may be computed slightly differently
  • Test validity concerns — in high-stakes assessment, having two unvalidated implementations undermines the psychometric defensibility of both

What Is NOT Duplicated#

Spaced repetition (SM-2, FSRS, HLR, Ebbinghaus, LECTOR, circadian-aware scheduling) is unique to Mnemosyne. Metis is focused on content generation and adaptive testing, not long-term memory scheduling.

Recommendation#

Create @shared/psychometrics exporting:

text
@shared/psychometrics/
  ├── irt/           # IRT models (1PL, 2PL, 3PL, MIRT)
  ├── cat/           # CAT engine, item selection strategies
  ├── item-bank/     # ItemBank with psychometric statistics
  ├── estimation/    # Ability estimation (MLE, EAP, MAP)
  └── information/   # Fisher information, test information functions

Both Metis and Mnemosyne consume this. Mnemosyne keeps spaced repetition as domain-specific.


4.2 AI Persona / Digital Human#

Severity: CRITICAL Domains: Psyche, Calliope, Hathor (Lilith excluded — no explicit persona system)

What Each Domain Implements#

Aspect Psyche Calliope Hathor
Service/Library Persona Service (port 8009) @calliope/genesis @hathor/llm-npc
Identity Model CRUD personas with personality config persona-architect.ts with psychological profile, artistic philosophy, biographical construction PersonalityManager with trait-based personality
Psychology Framework Voice/avatar personality configuration Artistic archetype, cultural grounding, genre identity via archetype-system.ts Big Five + Enneagram + BDI (Belief-Desire-Intention)
Emotional Model Not described Not described EmotionalStateManager — real-time emotional state tracking
Memory Not described Not described MemorySystem — short-term and long-term memory
Serialization PostgreSQL personas table identity-seed.ts (ArtisticIdentitySeed) Platform-agnostic (NVIDIA Ace, Inworld, Convai)

Cross-Domain Dependencies#

text
Psyche ──── (no dependency) ──── Calliope
   │                                │
   └──── (no dependency) ──── Hathor
                                    │
              Calliope ──── (no dependency) ──── Hathor

Zero cross-domain persona consumption. Three completely independent personality systems.

Why This Is Critical#

  • Scientific frameworks are shared — Big Five / OCEAN is a well-established model. Three different interpretations of the same psychological constructs will produce incompatible personality representations
  • No portability — an AI artist persona created in Calliope cannot appear as an NPC in Hathor or a digital human in Psyche, because each domain defines personality differently
  • Trait computation divergence — if Hathor implements Big Five scoring one way and a future Psyche update also adopts Big Five, they will score differently
  • Wasted effort — emotional state machines, personality evolution algorithms, and trait-to-behavior mappings are non-trivial to implement correctly

Recommendation#

Create @oshun/persona-core exporting:

text
@oshun/persona-core/
  ├── big-five/        # OCEAN trait model with scoring
  ├── enneagram/       # 9-type system with wings and growth/stress directions
  ├── emotions/        # Emotional state machine (PAD model or Plutchik's wheel)
  ├── bdi/             # Belief-Desire-Intention behavioral framework
  ├── memory/          # Short-term + long-term persona memory
  └── serialization/   # Persona export/import for cross-domain portability

Each domain extends with domain-specific layers:

Domain Extension
Psyche Adds voice profile, avatar configuration, real-time rendering parameters
Calliope Adds artistic philosophy, genre identity, cultural grounding, career arc
Hathor Adds NPC behavior trees, dialogue state, world knowledge, faction relationships

4.3 Voice Synthesis / TTS#

Severity: SIGNIFICANT Domains: Euterpe (canonical), Psyche (independent), Lilith (independent), Calliope (consumer), Iris (consumer), Metis (in progress)

Architecture Map#

text
                    ┌──────────────────────┐
                    │  EUTERPE (canonical)  │
                    │  @euterpe/voice       │
                    │  7 sub-modules:       │
                    │  - voice-cloning      │
                    │  - voice-conversion   │
                    │  - tts                │
                    │  - text-to-singing    │
                    │  - vocal-processing   │
                    │  - choir              │
                    │  - voice-analysis     │
                    └───────┬───────┬───────┘
                            │       │
                   consumes │       │ consumes
                            │       │
                    ┌───────▼──┐  ┌─▼────────┐
                    │ CALLIOPE │  │   IRIS    │
                    │ (bridge) │  │ (wrapper) │
                    └──────────┘  └──────────┘

    ┌────────────────────────┐    ┌────────────────────────┐
    │  PSYCHE (independent)  │    │  LILITH (independent)  │
    │  @psyche/voice-*       │    │  svc-tts               │
    │  5 libraries:          │    │  svc-voice-pipeline    │
    │  - voice-synthesis     │    │                        │
    │  - voice-streaming     │    │  Providers:            │
    │  - voice-consistency   │    │  - AWS Polly (primary) │
    │  - viseme-generator    │    │  - Chatterbox TTS      │
    │  - noise-handling      │    │                        │
    │                        │    │  Features:             │
    │  Providers:            │    │  - S3 audio caching    │
    │  - ElevenLabs (primary)│    │  - WebRTC SFU          │
    │  - XTTS v2             │    │  - Full-duplex audio   │
    │  - Cartesia            │    │  - Cross-device handoff│
    │  - OpenAI TTS          │    │                        │
    │  - Deepgram Aura       │    │  Does NOT consume      │
    │                        │    │  @euterpe/*            │
    │  Does NOT consume      │    └────────────────────────┘
    │  @euterpe/*            │
    └────────────────────────┘

Detailed Comparison#

Aspect Euterpe Psyche Lilith
Role Canonical voice domain Digital human voice Spiritual guidance voice
Architecture Pure library (7 sub-modules) 5 voice libraries 2 microservices
TTS Providers Custom AI/ML (diffusion, transformer, GAN) ElevenLabs, XTTS v2, Cartesia, OpenAI TTS, Deepgram Aura AWS Polly, Chatterbox TTS
Voice Cloning Yes (with consent gating) Yes (with fine-tuning) Not documented
Latency Target Not specified < 100ms p95 first-byte Not specified
Provider Switching Model backend swapping via infrastructure sub-module Dynamic provider-agnostic switching Primary + fallback only
Consumes Euterpe? N/A (is Euterpe) No No
Consumed By Calliope, Iris, Isis Iris (via @iris/integrations-psyche) Nothing
Unique Capability Singing synthesis, choir, vocal processing Viseme generation (lip sync), noise handling Full-duplex WebRTC, cross-device handoff

What's Duplicated#

The underlying TTS provider abstraction is fundamentally the same across all three:

  1. Accept text input + voice profile
  2. Call a provider API (ElevenLabs / OpenAI / AWS Polly / etc.)
  3. Receive audio buffer
  4. Stream to consumer with latency optimization
  5. Handle provider failures with fallback

Each domain implements this pattern independently.

What's NOT Duplicated#

  • Euterpe's singing synthesis and choir generation — unique to music production
  • Psyche's viseme generation — lip sync for photorealistic digital humans
  • Lilith's full-duplex WebRTC — real-time bidirectional voice chat
  • Metis — still unported from Minerva; not yet duplicating

Recommendation#

Create @oshun/tts-providers as a shared provider abstraction:

text
@oshun/tts-providers/
  ├── providers/
  │   ├── elevenlabs.ts
  │   ├── openai-tts.ts
  │   ├── aws-polly.ts
  │   ├── xtts.ts
  │   ├── cartesia.ts
  │   ├── deepgram-aura.ts
  │   └── chatterbox.ts
  ├── fallback-chain.ts     # Provider failover logic
  ├── audio-format.ts       # Format normalization (PCM, opus, mp3)
  ├── streaming.ts          # Chunked audio streaming
  └── types.ts              # VoiceProfile, TTSRequest, TTSResult

Psyche, Lilith, and Euterpe each consume this for raw TTS calls while maintaining their domain-specific orchestration on top (Psyche: visemes + lip sync; Lilith: WebRTC + caching; Euterpe: singing + choir + cloning).


4.4 Blockchain / Web3#

Severity: SIGNIFICANT Domains: Aje (canonical), Lilith (independent), Themis (opaque), Euterpe (planned only)

Architecture Map#

text
              ┌──────────────────────────────────────┐
              │          AJE (canonical)              │
              │  30+ libraries organized in layers:   │
              │                                       │
              │  Wallet:  @aje/wallets,               │
              │           @aje/account-abstraction     │
              │  DeFi:    @aje/defi, @aje/payments,   │
              │           @aje/oracles                 │
              │  Asset:   @aje/nft, @aje/rwa,         │
              │           @aje/storage                 │
              │  Protocol:@aje/contracts,              │
              │           @aje/bridges,                │
              │           @aje/governance,             │
              │           @aje/identity                │
              │  SDK:     @aje/sdk (unified entry)     │
              └───────────────┬────────────────────────┘
                              │
                   ┌──────────┼──────────────┐
                   │          │              │
                   ▼          ▼              ▼
            ┌──────────┐ ┌────────┐  ┌───────────┐
            │  THEMIS  │ │EUTERPE │  │  LILITH   │
            │ declared │ │planned │  │ BYPASSES  │
            │ but      │ │ only   │  │ @aje/*    │
            │ opaque   │ │        │  │ entirely  │
            └──────────┘ └────────┘  └───────────┘

Lilith's Independent Blockchain Stack#

Lilith has 14 dedicated blockchain microservices that bypass @aje/* entirely:

Lilith Service Equivalent in Aje
svc-blockchain (uses ethers.js 6 directly) @aje/core (crypto primitives)
svc-native-token @aje/contracts
svc-token-core @aje/contracts
svc-token-access @aje/contracts
svc-token-verification @aje/identity
svc-staking-mechanism @aje/restaking
svc-dao-governance @aje/governance
svc-defi-integration @aje/defi
svc-cross-chain-bridge @aje/bridges
svc-ipfs-integration @aje/storage
svc-micro-transaction @aje/payments
svc-fiat-ramp @aje/payments
svc-settlement @aje/payments

Key fact: Lilith uses ethers.js 6 directly for blockchain interaction rather than consuming any @aje/* library.

Themis Status#

Themis declares in its cross-domain integration table:

@themis/integrations — Governance decisions executed on-chain; on-chain data ingested into Themis

However, no @aje/* library appears in Themis's actual dependency tree. The integration is declared but the implementation path is opaque.

Euterpe Status#

Euterpe's blockchain music infrastructure is marked as (planned) only. No current blockchain code exists.

Recommendation#

Lilith's 14 blockchain services should be refactored to consume @aje/*:

Lilith Service Should Consume
svc-blockchain @aje/core (replaces raw ethers.js)
svc-native-token, svc-token-core, svc-token-access @aje/contracts
svc-staking-mechanism @aje/restaking
svc-dao-governance @aje/governance
svc-defi-integration @aje/defi
svc-cross-chain-bridge @aje/bridges
svc-ipfs-integration @aje/storage
svc-micro-transaction, svc-fiat-ramp, svc-settlement @aje/payments

Themis's @themis/integrations should explicitly document whether it wraps @aje/* or also uses ethers.js directly.


4.5 RAG / Knowledge Retrieval / Document Processing#

Severity: MODERATE Domains: Sophia (canonical), Iris (proper consumer), Nous (foundational), Lilith (independent), Veritas (independent)

Technology Comparison#

Aspect Sophia Iris Nous Lilith Veritas
Role Canonical knowledge domain AI assistant with KB Local inference backbone Consciousness platform News agency
RAG Libraries @sophia/indexing, @sophia/vectordb, @sophia/evaluation @iris/knowledge-* (8 libraries) @nous/llm rag/ subdirectory svc-rag, svc-knowledge, svc-ingestion @veritas/rag, @veritas/ingestion-core
Vector DB Qdrant, Pinecone, Milvus, Memory Qdrant + pgvector None (library-only) Milvus pgvector
Embeddings OpenAI, Cohere, Local Via @nous/llm Multi-provider manager Domain-specific spiritual embeddings Cohere
Chunking 8 strategies in @sophia/indexing @iris/knowledge-chunking Not provided Internal to svc-ingestion Internal pipeline
Graph DB N/A @iris/knowledge-graph N/A Neo4j N/A
Depends on Sophia? N/A Yes (@iris/integrations-sophia) No No No
Depends on Nous? No Yes (@nous/core, @nous/llm) N/A No No

Dependency Diagram#

text
                  ┌──────────────────┐
                  │  SOPHIA          │
                  │  (canonical RAG) │
                  │  8 chunking      │
                  │  strategies      │
                  │  4 vector DBs    │
                  └────────┬─────────┘
                           │ consumed by
                           ▼
                  ┌──────────────────┐     ┌──────────────────┐
                  │  IRIS            │────▶│  NOUS            │
                  │  (proper         │     │  (foundational   │
                  │   consumer)      │     │   embeddings)    │
                  └──────────────────┘     └──────────────────┘

    ┌──────────────────┐              ┌──────────────────┐
    │  LILITH           │              │  VERITAS          │
    │  (independent)    │              │  (independent)    │
    │  Milvus + Neo4j   │              │  pgvector + ES    │
    │  Spiritual embeds │              │  Cohere embeds    │
    │                   │              │                   │
    │  NO Sophia import │              │  NO Sophia import │
    │  NO Nous import   │              │  NO Nous import   │
    └──────────────────┘              └──────────────────┘

What Is Justifiably Different#

  • Lilith's spiritual embeddings — domain-tuned for consciousness/wellness content; generic embeddings don't capture the semantic space correctly
  • Veritas's pgvector choice — deliberately avoids operational complexity of a separate vector database for their scale (hundreds of thousands of articles)
  • Lilith's Neo4j — knowledge graph with ontologies and semantic relations specific to spiritual traditions

What IS Unnecessarily Duplicated#

  • Document chunking logic — all domains implement chunking independently when @sophia/indexing already has 8 strategies
  • Embedding provider abstraction — calling OpenAI/Cohere embedding APIs is identical across domains
  • Retrieval scoring and ranking — RRF fusion, similarity thresholds, re-ranking

Recommendation#

  1. Lilith and Veritas should consume @sophia/indexing for chunking strategies (even if they use different vector DBs)
  2. Lilith and Veritas should consume @nous/core embedding manager for provider abstraction (even if Lilith fine-tunes embeddings afterward)
  3. Vector DB choice remains domain-specific — Qdrant vs. Milvus vs. pgvector is a legitimate architectural decision

4.6 Meditation / Wellness#

Severity: MINOR / JUSTIFIED Domains: Meditation (canonical engine), Tara (clean consumer), Lilith (independent), Shakti (independent)

Architecture Map#

text
    ┌────────────────────────────┐
    │  MEDITATION (canonical)    │
    │  6 shared libraries:       │
    │  - meditation-player       │
    │  - meditation-timer        │
    │  - meditation-session      │
    │  - meditation-breathing    │
    │  - meditation-progress     │
    │  - meditation-offline      │
    └──────────────┬─────────────┘
                   │ consumes all 6
                   ▼
    ┌────────────────────────────┐
    │  TARA (clean consumer)     │
    │  Web (Next.js) + Mobile    │
    │  (Expo) + Admin Dashboard  │
    └────────────────────────────┘

    ┌────────────────────────────┐   ┌────────────────────────────┐
    │  LILITH (independent)      │   │  SHAKTI (independent)      │
    │  10 meditation services:   │   │  @shakti/yoga contains:    │
    │  - svc-meditation-core     │   │  - meditation.ts           │
    │  - svc-meditation-experience│   │    (11 meditation types:   │
    │  - svc-meditation-generation│   │     dharana, dhyana,       │
    │  - svc-group-meditation    │   │     pratyahara, yoga nidra, │
    │  - svc-breathwork          │   │     trataka, mantra, metta, │
    │  - svc-yoga-practice       │   │     chakra, savasana,      │
    │  - svc-spiritual-guidance  │   │     mindfulness,           │
    │  - svc-journal             │   │     visualization)         │
    │  - svc-notes               │   │  - pranayama.ts            │
    │  - svc-daily-content       │   │    (7 techniques: Ujjayi,  │
    │                            │   │     Nadi Shodhana,         │
    │  NO @meditation/* imports  │   │     Kapalabhati, etc.)     │
    │                            │   │  - ayurveda.ts (dosha)     │
    └────────────────────────────┘   │                            │
                                     │  NO @meditation/* imports  │
                                     └────────────────────────────┘

Why This Is Justified#

Lilith's independence makes sense because:

  • Lilith is a server-side microservices platform (80+ services) while libs/meditation/ are client-side engine libraries
  • Lilith's meditation integrates with AI generation (svc-meditation-generation), real-time group sync (svc-group-meditation), journaling (svc-journal), and content scheduling (svc-daily-content)
  • These are orchestration services that happen to involve meditation, not meditation engines

Shakti's independence makes sense because:

  • Shakti's pranayama includes Ayurvedic dosha effects (Ujjayi balances vata, Kapalabhati stimulates pitta, etc.)
  • Shakti's meditation types are yoga-lineage-specific (dharana, dhyana, pratyahara, trataka)
  • This is domain-specific knowledge modeling, not generic meditation timers

Minor Concern#

The basic breath timing model (inhale duration, hold duration, exhale duration, cycles) is reimplemented 3 times:

  • @oshun/meditation-breathing — 10 patterns
  • Lilith svc-breathwork — internal PatternService
  • @shakti/yoga pranayama.ts — 7 techniques

A tiny shared @oshun/breath-pattern type definition could reduce this without forcing architectural changes.


4.7 Movement / Motion Analysis#

Severity: ARCHITECTURE DOC INCONSISTENCY Domains: Aja (canonical), Shakti (proper consumer), Galatea (claims consumption but doesn't declare it)

Dependency Status#

Domain Role Depends on @aja/*? Evidence
Aja Canonical motion domain: pose estimation, 2D→3D lifting, motion quality, retargeting N/A (domain owner) N/A
Shakti Physical discipline platform: AI form analysis for yoga, strength, martial arts Yes — @shakti/form-analysis depends on @aja/pose-lifting Architecture doc explicitly states: "Shakti does not maintain its own pose estimation stack"
Galatea Humanoid robotics: perception, kinematics, choreography Claimed but not declared Architecture doc says "consumes motion science from Aja and Shakti" but the dependency graph section lists NO @aja/* or @shakti/* imports

The Inconsistency#

Galatea's architecture document contains two contradictory statements:

Statement 1 (Dependencies on Other Oshun Domains table):

@aja/* — Import motion capture data from Aja to teach robot motions by human demonstration @shakti/* — Human movement quality metrics inform robot motion aesthetics and biomechanical constraints

Statement 2 (Library Dependency Graph section): Lists only internal @galatea/* dependencies. No @aja/* or @shakti/* appears anywhere in the dependency tree.

Additionally, @galatea/perception builds its own complete computer vision stack (SLAM, depth processing, visual servoing, person detection) independently.

Recommendation#

Either:

  1. Add explicit dependencies: Add @aja/pose-lifting and @shakti/form-analysis to @galatea/perception or @galatea/ai dependency lists, OR
  2. Correct the documentation: Remove the Aja/Shakti claims from the cross-domain table if Galatea truly builds its own perception stack

4.8 Content Generation#

Severity: CLEAN (no action needed) Domains: Isis (canonical), Yemaya (proper consumer), Calliope (proper consumer)

Architecture — The Model Pattern#

text
    ┌────────────────────────────────────────────┐
    │  ISIS (canonical generative factory)        │
    │  15 generation modes, GPU workers,          │
    │  provenance tracking, workflow registry      │
    │                                              │
    │  Publishes: isis:jobs, isis:assets,          │
    │             isis:outputs streams              │
    └──────────────┬──────────────┬────────────────┘
                   │              │
          event bus│              │adapter pattern
                   │              │
    ┌──────────────▼──┐   ┌──────▼───────────────┐
    │  YEMAYA          │   │  CALLIOPE             │
    │  Publishes:      │   │  @calliope/bridge:    │
    │  yemaya.          │   │  - IsisAdapter        │
    │  generation.     │   │    .generateImage()   │
    │  requested       │   │    .trainLoRA()       │
    │                  │   │  - EuterpeAdapter     │
    │  Consumes:       │   │    .generateMusic()   │
    │  isis.job.       │   │  - AjaAdapter         │
    │  completed       │   │    .choreograph()     │
    └──────────────────┘   └──────────────────────┘

Why This Works#

  • Yemaya submits generation jobs via event bus (yemaya.generation.requested) and receives results via isis.job.completed. No generation engine code in Yemaya — only workflow orchestration
  • Calliope dispatches through explicit adapter interfaces in @calliope/bridge:
    • Image/video briefs → IsisAdapter
    • Music briefs → EuterpeAdapter
    • Motion briefs → AjaAdapter
  • Separation is clean: Isis handles compute, Yemaya handles production workflows, Calliope handles creative direction

This cluster should serve as the reference architecture for how cross-domain consumption should work throughout Oshun.


5. Cross-Domain Dependency Health#

Domains That Properly Consume Canonical Providers#

Consumer Provider Mechanism Quality
Calliope → Euterpe Voice/music generation @calliope/bridge EuterpeAdapter Excellent
Calliope → Isis Image/video generation @calliope/bridge IsisAdapter Excellent
Calliope → Aja Motion/choreography @calliope/bridge AjaAdapter Excellent
Iris → Nous LLM inference, embeddings @nous/core, @nous/llm Good
Iris → Sophia Research knowledge @iris/integrations-sophia Good
Iris → Euterpe Voice synthesis @euterpe/voice Good
Iris → Psyche Digital humans @iris/integrations-psyche Good
Shakti → Aja Pose estimation @aja/pose-lifting Good
Tara → Meditation All meditation engine All 6 @oshun/meditation-* libs Excellent
Yemaya → Isis Asset generation Event bus (yemaya.generation.requested) Excellent
Hathor → Sophia Citation grounding Event bus (sophia.document.ingested) Good

Domains That Should Consume But Don't#

Domain Should Consume Currently Uses Instead Impact
Psyche @euterpe/voice (or shared TTS) 5 independent TTS providers Significant
Lilith @aje/* for blockchain ethers.js 6 directly (14 services) Significant
Lilith @sophia/indexing for chunking Internal svc-ingestion Moderate
Lilith @nous/core for embeddings OpenAI/Anthropic APIs directly Moderate
Lilith @euterpe/voice (or shared TTS) AWS Polly + Chatterbox directly Moderate
Veritas @sophia/indexing for chunking Internal pipeline Moderate
Veritas @nous/core for embeddings Cohere directly Low
Mnemosyne Shared psychometrics Independent IRT/CAT implementation Critical
Metis Shared psychometrics Independent IRT/CAT implementation Critical
Calliope Shared persona core Independent persona architect Critical
Hathor Shared persona core Independent PersonalityManager Critical
Psyche Shared persona core Independent Persona Service Critical

6. Root Cause Analysis#

Why Does Duplication Exist?#

The duplication follows a consistent pattern:

Pattern 1: The "Large Domain" Effect#

Lilith (80+ microservices) is the single largest source of duplication. It builds its own:

  • TTS (AWS Polly + Chatterbox)
  • RAG (Milvus + Neo4j + spiritual embeddings)
  • Blockchain (14 services with ethers.js)
  • Meditation (10 services)

Likely cause: Lilith was developed as a monolithic product before the domain architecture was formalized. Its services were built to be self-contained, and the cost of retrofitting them to consume canonical domains was not prioritized.

Pattern 2: The "Different Provider" Justification#

Psyche uses ElevenLabs, Lilith uses AWS Polly, Euterpe uses custom AI/ML, Veritas uses Cohere for embeddings while Sophia uses OpenAI. Provider differences create the appearance of different systems when the abstraction layer is identical.

Likely cause: Each domain chose providers based on immediate needs without a shared provider abstraction layer existing first.

Pattern 3: The "No Shared Foundation" Gap#

Psychometrics (IRT/CAT) and persona modeling (Big Five/OCEAN) are well-established scientific frameworks that should have been extracted into shared libraries before domain development began. Without a shared foundation, each domain implemented from scratch.

Likely cause: These capabilities were needed by multiple domains simultaneously, and no coordination mechanism ensured a shared implementation was created first.

Pattern 4: The "Domain Purity" Overcorrection#

Some separation appears to be driven by a desire for domain independence rather than practical necessity. Veritas building its own RAG when Sophia has 8 chunking strategies is technically clean architecture (no coupling) but practically wasteful.


7. Recommendations#

Priority 1 — Critical (Create Shared Libraries)#

Action Effort Impact Affected Domains
Create @shared/psychometrics (IRT, CAT, ItemBank, estimation) Medium High — eliminates algorithmic divergence risk Metis, Mnemosyne
Create @oshun/persona-core (Big Five, Enneagram, BDI, emotions) Medium High — enables cross-domain persona portability Psyche, Calliope, Hathor

Priority 2 — Significant (Refactor to Consume Canonical Domains)#

Action Effort Impact Affected Domains
Migrate Lilith blockchain services to @aje/* High High — eliminates 14 redundant service implementations Lilith
Create @oshun/tts-providers shared abstraction Medium Medium — reduces provider management overhead Psyche, Lilith, Euterpe
Clarify Themis → Aje integration Low Low — documentation accuracy Themis

Priority 3 — Moderate (Share Abstractions Where Practical)#

Action Effort Impact Affected Domains
Lilith and Veritas consume @sophia/indexing for chunking Low Medium — avoids chunking logic duplication Lilith, Veritas
Lilith and Veritas consume @nous/core for embedding abstraction Low Low — reduces provider API boilerplate Lilith, Veritas

Priority 4 — Documentation Fixes#

Action Effort Impact Affected Domains
Fix Galatea architecture: either add @aja/* dependency or remove the claim Trivial Low — documentation accuracy Galatea
Area Reason
Lilith meditation services Justified: server-side orchestration with AI generation, group sync, journaling — fundamentally different from client-side engine libraries
Shakti yoga pranayama Justified: Ayurvedic dosha effects and yoga-lineage-specific meditation types are domain knowledge, not generic timers
Veritas pgvector choice Justified: deliberate architectural decision to avoid operational complexity of separate vector DB
Lilith Milvus + Neo4j choice Justified: spiritual domain-tuned embeddings and ontological knowledge graph require specialized backends

Appendix A — Full Domain Purpose Index#

# Domain Purpose Library Count
1 Aglaea AI fashion, beauty, personal styling ~150+ biomarkers
2 Airmid Evidence-based phytotherapy and botanical intelligence Pure library
3 Aja Motion AI and animation production 38 libraries
4 Aje Blockchain, DeFi, Web3 infrastructure 30+ libraries
5 Aphrodite Adult entertainment streaming platform Multiple libs
6 Arete Personal development and life mastery 42-table schema
7 Asase Food and agriculture ops for Ghana Multiple libs
8 Bellona Build bridges to game engines Multiple libs
9 Calliope Autonomous AI artist creation and management Orchestration domain
10 Contracts Cross-domain entity schemas and events Shared
11 Demeter Home gardening and small-scale agriculture Multiple libs
12 Euterpe Comprehensive music technology 7 libraries
13 Galatea Humanoid robotics for fashion retail 20 libraries
14 Hathor Worldbuilding, narrative design, simulation Multiple libs
15 Hestia Culinary intelligence and smart home Multiple libs
16 Iris Universal intelligent assistant platform 43 libraries
17 Isis Generative AI factory 35 libraries
18 Kuanyin Platform-wide ethics, safety, moderation Multiple libs
19 Lilith Conversational AI and consciousness experience 80+ services
20 Maat Autonomous organization OS for West Africa Multiple libs
21 Maya Metaverse / game engine 9 existing + ~120 planned
22 Meditation Platform-agnostic meditation engine 6 libraries
23 Metis AI-powered educational content platform Multiple libs
24 Mnemosyne Humanistic learning and cultural intelligence Multiple libs
25 Neith Sovereign Rust runtime kernel Rust crates
26 Nisaba Ancient text analysis (18 writing systems) Multiple libs
27 Nous Sovereign AI inference backbone Multiple libs
28 Nyx Cosmic observatory platform Multiple libs
29 OpenAPI Centralized OpenAPI spec management Shared
30 Oshun Platform shell and domain orchestration Shell domain
31 Oya Drone and autonomous aerial systems Foundational
32 Proto Protocol Buffer definitions for gRPC Shared
33 Psyche Hyper-realistic AI virtual assistant 131 libraries
34 Seshat Dwelling arts and craftsmanship intelligence Multiple libs
35 Shakti Physical discipline and movement intelligence 27 libraries
36 Shared Infrastructure foundation (30 libraries) 30 libraries
37 Sophia Knowledge management and research 24 libraries
38 Tara Meditation and mindfulness app 3 apps + 6 libs
39 Themis Governance and compliance infrastructure Multiple libs
40 Uzume Live entertainment production 22 libraries
41 Veritas Autonomous AI news agency for Ghana 63 libraries
42 Yemaya Creative production studio 41 libraries

Appendix B — Duplication Severity Matrix#

text
                    CRITICAL        SIGNIFICANT      MODERATE         MINOR/JUSTIFIED   CLEAN
                    ────────        ───────────      ────────         ───────────────   ─────
Psychometrics       Metis ←→
(IRT/CAT)           Mnemosyne

Persona             Psyche ←→
(Big Five/OCEAN)    Calliope ←→
                    Hathor

Voice/TTS                           Psyche,                                            Calliope
                                    Lilith                                             → Euterpe
                                    (vs. Euterpe)                                      Iris
                                                                                       → Euterpe

Blockchain/Web3                     Lilith
                                    (vs. Aje)

RAG/Knowledge                                        Lilith,                           Iris
                                                     Veritas                           → Sophia
                                                     (vs. Sophia)                      → Nous

Meditation                                                            Lilith,          Tara
                                                                      Shakti           → Meditation
                                                                      (justified)

Motion Analysis                                                                        Shakti
                                                                                       → Aja
                                                                      Galatea (doc
                                                                      inconsistency)

Content                                                                                Yemaya
Generation                                                                             → Isis
                                                                                       Calliope
                                                                                       → Isis

End of report.