Status: accepted · Date: 2026-07
Ollama + Llama 3.2 3B was installed and benchmarked as a potential local reasoning layer.
Rejected as primary: measured ~93 seconds for trivial inference on the Pi 4 CPU. Cloud AI (Claude Code) handles vault reasoning and organization; local processing is reserved for cases where data sensitivity outweighs everything else.
- The sensitive-content boundary must be enforced by workflow (which files get read into cloud context), not by tooling defaults. Documented in the vault's usage conventions.
- Revisit if the node is ever replaced with hardware that changes the latency math.