Raspberry Pi Lab

In June 2026 I set the Pi up as an always-on checker for the docs of Great Hearth, my household app. At two in the morning it looked for broken links, missing index entries and outdated terms, then opened a pull request with its report. I merged 13 of its reports after checking them with Claude Code, Anthropic's coding agent.

The Pi could not reach my Mac, so every two hours the Mac committed a copy of the docs to GitHub for the Pi to pull. In July I checked what the Pi was adding, and the numbers in the first point below made the case for moving the job.

On 2026-07-21 I moved the check to my Mac, which reads the live docs directly, and deleted the copy and the sync job that served only the Pi. Docs change over weeks, so the check runs when I ask for it.

• What the Pi's model did. The Pi also carried a language model meant to find contradictions between documents. At 1.5 billion parameters, a size that fit in the Pi's memory, the model was too weak for the task. All 13 reports the Pi archived skipped the contradiction step, and the model answered zero requests in the 30 days before 2026-07-21.

• The same step on the Mac. My current Mac runs a model of about 80 billion parameters. It drafted contradiction findings in 4 of the 5 reports since the handover, and Claude Code confirms a finding before I act on it.

• Two ideas that never ran on a Pi. A voicemail line I planned for the Pi has a requirements draft and no code. In June 2026 I weighed buying a new Pi for my coding model and kept it on the Mac I had then.