Skip to main content
Version: v0.6.0

Side-Chain Inference Pattern

The Side-Chain Inference pattern is the architectural standard for integrating Artificial Intelligence (ML/LLM) into the fluxrig data plane. It ensures that probabilistic models can provide high-value insights (scoring, categorization, anomaly detection) without compromising the Determinism or Performance of the primary execution path.

The Theory: Dry vs. Wet Signals​

  • The Dry Signal (Deterministic): The core transaction logic (e.g., "Is this ISO8583 message valid?") which must be 100% reproducible and low-latency.
  • The Wet Signal (Probabilistic): The AI-augmented insight (e.g., "What is the probability this is a fraudulent transaction?") which is non-blocking and descriptive.

Implementation Blueprint​

Integrating an AI model (e.g., via the Sovereign AI Bridge) follows a three-stage lifecycle: Tap, Infer, and Feed.

1. The Signal Tap (Aux Send)​

A dedicated Gear (often the Bento Gear) acts as an "Aux Send". It receives a clone of the fluxMsg from the main logic flow.

wires:
  - from: ingress.out
    to:   ai-bridge.in      # The Signal Tap
  - from: ingress.out
    to:   business-logic.in # The Hot Path

2. Isolated Inference (Sovereign Bridge)​

The Bridge Gear transmits the signal to a local inference engine (e.g., Ollama, TensorRT, or a specialized Sovereign AI node).

  • Out-of-Process: Inference happens outside the Rack's core execution loop to prevent CPU/Memory starvation.
  • Non-Blocking: The primary business logic continues to execute while the AI model is "thinking."

3. Metadata Feedback Loop​

Once the AI model completes its analysis, the Bridge Gear emits a Feedback Signal. Since the original message has already moved forward, the AI result is typically attached to the Signal Metadata of the next related signal or stored in a shared state (e.g., Coat Check).

// Example Metadata Feedback Structure
msg.Metadata["flux.ai.fraud_score"] = "0.92"
msg.Metadata["flux.ai.rationale"] = "unusual_temporal_cluster"

Security Sovereignty​

In the Side-Chain pattern, the raw sensitive payload (e.g., a credit card number) is often masked before being sent to the AI Bridge. The AI engine only receives the "Contextual Vectors" required for analysis, ensuring that PII never leaves the deterministic vault.

Technical reference​