AAA-shadow-mode

AAA-shadow-mode — Sovereign Shadow Multimodal Compute

1. Purpose & Doctrine

The Shadow Mode architecture establishes an out-of-band, high-performance compute plane dedicated to heavy multimodal AI generation (Image, Video, Audio) and sovereign human needs.

In accordance with the Zen Doctrine and F13 Sovereignty:

1. Machine Peace: The primary VPS remains lean and uncluttered. Heavy CUDA/PyTorch tensor computation is offloaded entirely to dedicated shadow instances (e.g., A100 80GB on Hostinger/RunPod).
2. Agent Peace: Jobs are dispatched asynchronously via REST API / worker queues with task receipts. Agents do not block or freeze waiting on generation.
3. Human Peace & Privacy: All personal creative, artistic, and exploratory tasks execute in secure, sovereign isolation with zero external telemetry leakage.

┌────────────────────────────────────────────────────────┐
│                   PRIMARY VPS (Lean)                  │
│   arifOS (8088) ── arifFLOW (7073) ── HERMES / AAA    │
└──────────────────────────┬─────────────────────────────┘
                           │ Async Job Dispatch (SSH / REST)
                           ▼
┌────────────────────────────────────────────────────────┐
│               SHADOW GPU INSTANCE (Heavy)              │
│   • NVIDIA A100 (80GB VRAM) / Headless Daemon          │
│   • ComfyUI REST API (:8188)                           │
│   • Flux.1-dev/schnell, SDXL, Wan2.1, CosyVoice        │
│   • Isolated Output Cache (/home/ubuntu/ComfyUI/output)│
└────────────────────────────────────────────────────────┘

2. The Three Invariants of Shadow Mode

Invariant 1: Physical & Process Isolation (F1 AMANAH)

• Never run multi-gigabyte diffusion/video model inference directly on the primary VPS.
• All heavy weights (.safetensors, .gguf, checkpoints, UNETs) live on the shadow instance storage.
• Network access between the VPS and Shadow instance is authenticated via dedicated Ed25519 keys (/root/.ssh/hostinger_ed25519).

Invariant 2: Sovereign Discretion & Human Needs (F13 SOVEREIGN)

• The shadow plane exists to fulfill the sovereign operator's creative, artistic, visual, and personal needs without external SaaS censorship, throttling, or telemetry snooping.
• Generative tasks respect legal boundaries, fictional character domains, and sovereign consent.
• Generated artifacts are stored in private directories with explicit lifecycle management (retain or auto-purge).

Invariant 3: Asynchronous EMD Contract (Encode → Metabolize → Decode)

• Encode: Hermes/AAA captures the user prompt, style constraints, resolution, and seed, encoding them into a standardized JSON workflow.
• Metabolize: The task is queued out-of-band. The primary agent returns immediate confirmation with a Task ID / Receipt.
• Decode: The shadow GPU renders the batch, posts back completion, and provides the output URL or directly streams the final media artifact.

3. Supported Multimodal Capabilities

4. Operational Runbook

A. Health & State Check

To verify shadow compute availability from the primary VPS:

# Check SSH connectivity
ssh shadow-gpu "nvidia-smi --query-gpu=name,memory.used,memory.total,utilization.gpu --format=csv,noheader"

# Check ComfyUI API health
ssh shadow-gpu "curl -s http://127.0.0.1:8188/system_stats"

B. Job Submission Pattern (ComfyUI REST API)

1. Prepare the JSON workflow payload with node graph mappings.
2. Dispatch payload to http://127.0.0.1:8188/prompt.
3. Track execution state via /history/{prompt_id}.
4. Retrieve the resulting image/video from /view?filename={filename}&type=output.

C. Resource Zen (Zero Drift)

• Checkpoints stored in /home/ubuntu/ComfyUI/models/checkpoints/.
• LoRAs stored in /home/ubuntu/ComfyUI/models/loras/.
• Output artifacts cached in /home/ubuntu/ComfyUI/output/.
• When idle, VRAM is released or cached efficiently without consuming primary VPS memory.

5. Failure Modes & Self-Healing

6. Audit & Receipts

Every shadow-mode dispatch records a receipt to VAULT999:

• task_id: Unique generation identifier.
• model_used: Checkpoint / LoRA / VAE fingerprint.
• vram_allocated: VRAM consumed during run.
• duration_ms: Total execution time.
• status: SUCCESS / FAILED / ROLLBACK.