Design

Design Principles

WWHAD

What would Home Assistant Do? Home Assistant has battle hardened patterns for complex configurations and intuitive UIs to set those configurations.

Human readable

Every "decision" should leave a human readable audit trail.
Note: I put "decision" in quotation marks because LLMs do not make decisions, they make next token predictions.

Default to code

Asking an LLM to do a programmable task is the robot equivalent of this meeting could have been an email. Don't burn tokens trying to convince a non deterministic machine to do a programmable task.

Zero trust

Patterns should not rely on trusting a non deterministic machine.

Optimize for memory

Memory is most valuable for inference. The disk based queueing system was added for human readability and to reduce memory usage.

Provenance

Not all classification is equal. An email classified by an LLM is a probabilistic guess. An email classified by its domain name is deterministic. The system tracks this distinction as provenance (rule, llm, or hybrid) and uses it to gate irreversible actions. Automations with LLM provenance cannot trigger irreversible actions unless explicitly overridden with !yolo.