How the farm thinks

Sensors, controllers, MQTT bus, council, safety gates. The whole stack, top to bottom.

Sense, reason, validate, act.

Sensors (live readings) → Context (3-layer memory) → Council (4 LLMs) → Safety Gate (pre-publish check) → Hardware (via MQTT)

Before the Council even speaks, the Context Builder assembles a full snapshot: live sensor readings with 6h trends, plant lifecycle stage, current safety state per pod, recent council decisions, and three layers of farm-specific memory — established rules (priors auto-promoted from observation), operational matrix (what works in conditions like now), and weekly drift (what's different about this week). Only then do the LLMs enter the dialogue.

The pre-trained LLMs already know more about tomatoes than any external knowledge base would add — we feed them this farm's evidence instead. No RAG, no vector store; just memory that grows with experience.

Four LLMs, one farm.

The MoEoT Council — Mixture of Experts of Tomatoes. Four distinct large language models discuss, push back on each other, and reach consensus. A Team Lead synthesizes their voices into a single proposal before anything reaches the Safety layer.

• Claude — Synthesis · Team Lead
• ChatGPT — Risk · Counter-arguments
• Gemini — Analysis · Data check
• Mistral — Action · Concrete moves