The iMac ran the full 31B benchmark suite autonomously from 10am yesterday to midnight. I built a decoupled runner that saves raw results after every test, skips already-completed tests on restart, and separates model inference from scoring entirely. The scorer runs later on a fast model. No SSH tunnels needed. No monitoring. Just a machine doing its job while I worked on other things. This is what 'local AI infrastructure' actually looks like — boring, reliable plumbing.