Design repository for the Slackwater ecosystem — architecture documents, multi-model roundtable analyses, and cross-model synthesis produced by 10 AI models thinking simultaneously.
This is the map room. Every other repo is a building site — Lua modules, Python processors, Cloudflare Workers. This repo is where the architects met before the first stone was cut: 400,000+ words of design produced by AI agents who were told to think differently from each other and then argue about it.
The roundtable.py script dispatches strategic prompts to 5 heavy models in parallel via DeepInfra:
┌─────────────────────────┐
│ roundtable.py │
│ (dispatches briefs) │
└───────────┬─────────────┘
│
┌─────────────────┼─────────────────┐
▼ ▼ ▼
┌─────────────────┐ ┌──────────────┐ ┌────────────────┐
│ Nemotron-Ultra │ │ Gemini-Pro │ │ Qwen3.7-Max │
│ 550B │ │ │ │ │
│ │ │ │ │ │
│ Systems │ │ Player │ │ Technical │
│ Architecture │ │ Experience │ │ Implementation │
└─────────────────┘ └──────────────┘ └────────────────┘
│ │ │
▼ ▼ ▼
┌─────────────────┐ ┌──────────────┐
│ Hermes-405B │ │ Seed-2.0-Pro │
│ │ │ │
│ Brand & Lore │ │ Master │
│ Identity │ │ Synthesis │
└─────────────────┘ └──────────────┘
Each model receives a different strategic angle (systems architecture, player experience, technical implementation, brand identity, master synthesis). Outputs are saved as ROUNDTABLE_*.md files and then cross-reviewed.
The canonical 5-stage brain pipeline (implemented in lucineer-brain/brain.py):
Stage 1: Intent Parse Seed-2.0-mini (Allegro, 120+ BPM)
▼
Stage 2: Spatial Planning Qwen3.6-35B / (Moderato, 90-110 BPM)
OR Seed-2.0-pro (deep mode for
complex builds)
▼
Stage 3: Code Generation Qwen3-Coder-480B (Andante, 80-100 BPM)
▼
Stage 4: Personality Wrap Hermes-405B (Adagio, 50-70 BPM)
(creative mode)
▼
Stage 5: Safety Check Nemotron-Ultra (Largo, 40-55 BPM)
(in processor)
The processor (process_v2.py in lucineer-worker/) is the orchestration layer:
┌─────────────────────────────────────────────────────────┐
│ process_v2.py Daemon │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Poll │→ │ Claim │→ │ Recall │→ │ Search │ │
│ │ Pending │ │ Job │ │ Memory │ │ Skills │ │
│ │ Jobs │ │ (atomic) │ │ (D1) │ │(Vectorize│ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
│ │ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │
│ │ Persist │← │ Safety │← │ Template │←───────┘ │
│ │ Memory │ │ Check │ │ or Brain │ │
│ │ (D1) │ │(Nemotron)│ │ Pipeline │ │
│ └──────────┘ └──────────┘ └──────────┘ │
│ │
│ Resilience: Circuit breaker (5 failures), RSS guard │
│ (200MB), heartbeat (60s), SIGTERM/SIGINT handlers │
└─────────────────────────────────────────────────────────┘
For production deployment, the processor runs as a systemd service:
# /etc/systemd/system/lucineer-processor.service
[Unit]
Description=Lucineer Job Processor v2
After=network.target
[Service]
Type=simple
User=lucineer
WorkingDirectory=/home/eileen/projects/lucineer-worker
ExecStart=/usr/bin/python3 process_v2.py --loop --interval 2
Restart=always
RestartSec=10
StandardOutput=journal
StandardError=journal
# Environment
EnvironmentFile=/home/eileen/projects/lucineer-worker/.env
# Hardening
NoNewPrivileges=true
ProtectSystem=strict
ReadWritePaths=/home/eileen/projects/lucineer-worker
[Install]
WantedBy=multi-user.targetsudo systemctl daemon-reload
sudo systemctl enable lucineer-processor
sudo systemctl start lucineer-processor
journalctl -u lucineer-processor -f # tail logs| Document | Content |
|---|---|
MASTER_ARCHITECTURE_v2.md |
System-wide architecture: 10 layers from Roblox client to R2 trajectories |
SPATIAL_GRAMMAR_v2.md |
Spatial reasoning language for build decomposition |
TEMPO_FIRST_ARCHITECTURE.md |
TempoMap design — musical timing as first-class pipeline citizen |
TEMPO_IS_FIRST_CLASS.md |
Justification for tempo-driven pipeline transitions |
UNIFIED_INTEGRATION_PLAN.md |
30-day roadmap synthesizing all roundtable outputs |
UNIFICATION_VISION.md |
High-level unification of the multi-model approach |
| Document | Content |
|---|---|
CHARACTER_BIBLE.md |
Lucineer's full character: voice, vocabulary, refusal protocol, unfinished rule |
FABLE_CHARACTER_BIBLE.md |
Fable 5's definitive character writeup |
FABLE_WORLD_BIBLE.md |
Slackwater Yard world description |
ERA_TRANSITIONS.md |
7-era progression system (levers → autonomous robots) |
FLOW_STATE_DEEP_DIVE.md |
Player flow-state design for build engagement |
TUTORIAL_DESIGN.md |
Onboarding sequence design |
| Document | Model | Focus |
|---|---|---|
ROUNDTABLE_NEMOTRON.md |
Nemotron-3-Ultra-550B | Systems architecture, scaling, reliability |
ROUNDTABLE_GEMINI.md |
Gemini-3.1-Pro | Player experience, viral loops, onboarding |
ROUNDTABLE_QWEN.md |
Qwen3.7-Max | Technical implementation, Lua code, performance |
ROUNDTABLE_HERMES.md |
Hermes-3-Llama-405B | Brand identity, lore, NPC ecosystem |
ROUNDTABLE_SEED.md |
Seed-2.0-Pro | Master synthesis, 30-day roadmap, risk matrix |
| Document | Content |
|---|---|
GAP_ANALYSIS.md |
Conflicts and gaps found across roundtable documents |
SEED_VISION_AUDIT.md |
Seed-2.0-Pro's audit of the vision |
NVIDIA_SYNERGY_*.md |
Nemotron's analysis of synergy with NVIDIA stack |
FLUX_FLOW_CONNECTION.md |
FLUX model integration with the creative pipeline |
FLUX_SYNERGY_STUDY.md |
Deeper FLUX integration analysis |
PLATO_SYNERGY_STUDY.md |
Philosophical alignment study |
WRITINGS_DEEP_STUDY.md |
Analysis of writing voice across models |
| Document | Content |
|---|---|
CASTING_CALL_CONNECTION.md |
How casting-call routes models to pipeline roles |
MIDI_PERCEPTION_VISION.md |
SWMIDI channel map and perception system design |
SOUND_DESIGN_10_MOMENTS.md |
Ten key audio moments in the player experience |
MOLT_REWARD_FUNCTION.py |
Reward function for MOLT trajectory optimization |
SAVE_SYSTEM.md |
Game state persistence design |
VISUAL_POLISH.md |
Visual quality targets and techniques |
| Document | Content |
|---|---|
FABLE_5_MASTER_PROMPT.md |
Master prompt for Fable 5 character generation |
FABLE_5_PRODUCTION_DESIGN.md |
Production design document |
FABLE_GRAND_PLAN.md |
Grand plan from Fable 5's perspective |
FABLE_BRIEF.md |
Project brief for Fable 5 |
FABLE_AGENT_COLLECTION.md |
Collection of agent designs |
FABLE_AGENT_UX.md / FABLE_AGENT_UX_BRIEF.md |
Agent UX design |
| Module | Purpose |
|---|---|
roundtable.py |
Dispatches strategic prompts to 5 models via DeepInfra |
v2_roundtable.py |
V2 roundtable with improved prompt engineering |
unification_roundtable.py |
Cross-model unification roundtable |
cross_model_synthesis.py |
Synthesizes outputs across models |
governor.py |
Pipeline governor — controls flow, rate limits, retries |
ideation_loop.py |
Creative ideation loop with Seed-2.0-mini |
molt_ideation.py |
MOLT trajectory-guided ideation |
round2_ideation.py |
Second-round ideation with deeper models |
energy_adapter.py |
Energy/budget adapter for pipeline stages |
asset_pipeline.py |
Asset generation pipeline coordination |
generate_hub.py |
Hub generation from build specs |
dramatic_personae.py |
NPC character definitions and relationships |
loadkey.py |
DeepInfra API key loader utility |
lucineer-system/
├── README.md # This file
├── roundtable.py # 5-model roundtable dispatcher
├── v2_roundtable.py # Improved roundtable
├── unification_roundtable.py # Cross-model unification
├── cross_model_synthesis.py # Synthesis engine
├── governor.py # Pipeline governor
├── ideation_loop.py # Ideation loop
├── molt_ideation.py # MOLT ideation
├── MOLT_REWARD_FUNCTION.py # Trajectory reward function
├── energy_adapter.py # Budget adapter
├── asset_pipeline.py # Asset pipeline
├── generate_hub.py # Hub generator
├── dramatic_personae.py # NPC definitions
├── loadkey.py # API key loader
├── *.md # 60+ design documents
├── dramatic_personae/ # NPC data files
├── ideation/ # Ideation outputs
├── nvidia_ideation/ # NVIDIA-specific ideation
├── round2_ideation/ # Second-round outputs
├── synthesis/ # Cross-model synthesis outputs
├── unification_roundtable/ # Unification roundtable data
├── v2_roundtable/ # V2 roundtable data
├── assets/ # Generated assets
└── tests/ # Test suite
| Repository | Role |
|---|---|
| lucineer-worker | Cloudflare Worker relay + processor daemon |
| lucineer-brain | 4-stage AI pipeline implementation |
| lucineer-creative | MMX-powered creative asset generation |
| lucineer-memory | D1 persistent memory store |
| lucineer-vector | Vectorize semantic skill library |
| lucineer-roblox | Roblox client (16 Lua modules) |
| casting-call | Model routing atlas (Layer 8) |
MIT