An AI-run home lab

Most home labs are a pile of things that work. This one is an argument about how software should get built when the person building it is one man with evenings.

The rule is that AI agents write nearly all of it, and nothing they write is trusted on the strength of looking right. Every change runs a suite, and the suites get audited in their turn: broken on purpose to check they can still go red, because a test that has never failed is decoration. The last full sweep of that turned up plenty of tests that couldn’t, which is rather the point of measuring it instead of assuming. Guard hooks refuse edits that would land a scheduled job with no failure alert, or shell code the linter rejects, at the keystroke rather than in review. Continuous integration re-runs the lot on a different machine, because “it worked on mine” is not evidence.

What comes out is less interesting than what it demonstrates: an environment that catches its own mistakes faster than I could read them.