The Five-Stage Pipeline

Every input that reaches the Neural Core — a sensor reading, a query, an event from a connected system — passes through the same five-stage pipeline. Each stage is designed for deterministic operation with bounded latency under load, so operators can plan infrastructure capacity with confidence.

  • 1. Input Normalization: Raw data from diverse sources — physical sensors, digital data streams, internal systems — is converted into a unified internal format before any content processing begins.
  • 2. Context Retrieval: The system queries its own memory: has this pattern occurred before? What history is relevant? This step connects the current input to the consciousness layer.
  • 3. Inference: The actual processing — classification, evaluation, decision-making — runs locally on the available hardware, with no cloud offload.
  • 4. Output Synthesis: The result is translated into an action or response — an autonomous reaction, a flag for operator review, or both.
  • 5. Memory Consolidation: What was learned in this cycle is written into the consciousness layer's five-database memory architecture, rather than discarded once the request completes.

Example: Anomaly Detection

Henri's anomaly detection — one of the system's nine independent detection skills — runs a related three-stage classification pipeline: signal fingerprinting of sub-threshold deviations, cross-cycle correlation of recurring patterns, and finally classification with either autonomous action or a flag for operator review. In the current production deployment, classification latency runs under 40 milliseconds per cycle — entirely on embedded hardware, with no dedicated GPU infrastructure.

Pipeline Stages5 (Normalization to Consolidation)
Operating ModeDeterministic, Bounded Latency
Observed Classification Latency< 40ms per Cycle (Production Deployment)
Compute LocationFully Local, No Cloud Offload