AI can compute. Trinity is for the imagination.
Trinity is a prompting system that prompts both the human and the AI. It uses Socratic questioning to draw out the human's vision before any work begins, then executes through a 12-phase instructional design methodology (ADDIECRAPEYE). The result: better output, because the AI understands what you actually want before it starts building.
Runs on a 128 GB AMD Strix Halo (Zen 5 + RDNA 3.5 + XDNA 2 NPU) with local LLMs and ComfyUI creative pipeline — no cloud APIs, no per-token costs, everything on-device.
| Layer | What it does | How |
|---|---|---|
| Socratic (Depth) | AI treats user as SME, questions before working | P (DiffusionGemma) or H (Hermes) |
| FACES (Image) | Emotional context via 4-byte protocol | NPU-accelerated (moved to Semantic Slime project) |
| LitRPG (Narrative) | Lightweight quest scaffolding for workflows | ADDIECRAPEYE state machine |
The pipeline: Human Imagination → Socratic questioning → FACES (emotional context) → LitRPG (quest framing) → ADDIECRAPEYE (12-phase execution) → Strix Halo (GPU: LLM, NPU: FACES) → Android XR (imagination visible in space)
| Device | Role | Hardware |
|---|---|---|
| Phone (Director) | Human input, Socratic questioning, quest management | Pixel 10 Pro XL, Gemini Nano |
| Desktop (Engine) | AI inference, orchestration, content generation | Strix Halo, 128GB VRAM |
| XR (Canvas) | Spatial review, EYE phase | XREAL Aura (Fall 2026) |
Authority flow: Phone (commands) → Desktop (compute) → XR (spatial output)
- Linux system with 16 GB+ RAM (developed on 128 GB Strix Halo)
- vLLM on port :8000 (P: DiffusionGemma 26B — execution engine)
- ComfyUI on port :8188 (A: creative pipeline — image, voice, video, 3D)
- Rust 1.80+
# 1. Start P (DiffusionGemma 26B) — always-on via podman/vLLM on :8000
bash ~/trinity-models/start-diffusiongemma.sh
# 2. Start Trinity server
cargo run -p trinity -- --headless
# 3. Launch A (ComfyUI) via hotel manager
curl -X POST http://localhost:3000/api/inference/hotel/studio
# 4. Open on phone or desktop:
# http://100.83.222.35:3000/trinity/phone.html (Tailscale)
# http://localhost:3000/trinity/phone.html (local)| Role | Model | Engine | Port | VRAM | Purpose |
|---|---|---|---|---|---|
| P | DiffusionGemma 26B AWQ-INT4 | vLLM (podman) | :8000 | ~42GB | Execution — story, code, tool calling, reasoning |
| A | ComfyUI + Janus-Pro-7B + VibeVoice-1.5B | ComfyUI | :8188 | ~17GB | Creative — image gen, voice, video, 3D on demand |
Both are always resident in Studio mode (~59GB of 128GB). Tier 2 models (HunyuanVideo, TRELLIS, LongCat, FLUX, TripoSR) are loaded on demand by ComfyUI.
Hermes 4 70B (H): Managed externally by LM Studio on :8002 when needed for planning. Submits jobs to Trinity via POST /api/jobs.
| Feature | Description |
|---|---|
| Chat Mode | DiffusionGemma 26B executes tasks with tool calling, Socratic questioning |
| Phone PWA | Chat, voice input, TTS, hotel controls, image gen, focus modes, RAG, persistent memory |
| Hotel Manager | 3 modes: Studio (P+A), Solo (P only), Closed (all off) |
| RAG | Semantic search via in-process ONNX Runtime embeddings (nomic-embed-text INT8) |
| Persistent Memory | SQLite-backed cross-session memory, injected into agent system prompt |
| ADDIECRAPEYE | 12-phase instructional design state machine with Bloom's Taxonomy mapping |
| FACES Protocol | 4-byte emotional context protocol (moved to Semantic Slime project July 2026) |
| EYE Package Export | HTML5 quiz games, text adventures, DOCX documents |
| Security | Bearer token auth on dangerous endpoints, rate limiting, input validation |
TRINITYIDAIOS/
├── crates/
│ ├── trinity/ # Main Axum server, agent loop, tools, inference
│ ├── trinity-protocol/ # Shared types (CharacterSheet, QuestState, etc.)
│ ├── trinity-quest/ # Quest engine (ADDIECRAPEYE state machine)
│ ├── trinity-iron-road/ # Book writing, VAAM vocabulary
│ ├── trinity-voice/ # Voice pipeline
│ ├── trinity-daydream/ # Bevy 3D/XR (future)
│ └── trinity-mcp-server/ # MCP server for extensibility
├── docs/active/ # Current docs (MASTER_PIVOT, MASTER_TASK_LIST, SECURITY, etc.)
├── docs/archive/ # Historical docs (preserved, not active)
├── scripts/launch/ # trinity_day.sh, trinity-start.sh — startup scripts
├── scripts/focus/ # creative-focus.sh, code-focus.sh, night-shift.sh
├── context.md # System context (kept current)
└── README.md # This file
| Doc | Purpose |
|---|---|
docs/active/MASTER_PIVOT_DOCUMENT.md |
Core vision, three-device architecture, honest assessment |
docs/active/MASTER_TASK_LIST.md |
Architecture audit, task tracking, completion criteria |
docs/active/ADDIECRAPEYE_CANONICAL.md |
12-phase instructional design reference |
docs/active/SECURITY.md |
Security model for public deployment |
docs/TRINITY_IDENTITY.md |
Authoritative 3-tier architecture: Core, Middleware, Products |
context.md |
Current system state, port map, what's working |
FACES is a 4-byte emotional context protocol designed to be the TCP/IP of emotive AI. It encodes Focus, Action, Container, Energy, and State into a single u32.
Status: Moved to the Semantic Slime project on July 5, 2026. W1-W6 complete (283 tests + 15 doc-tests). The trinity-faces crate no longer exists in this workspace. Trinity's Triple Reflection philosophy still references FACES as the "Image" layer, but the implementation is an external product dependency.
Private project. Public deployment at LDTAtkinson.com for educational demonstration.