Local AI memory locked in encrypted vault, 9 tests run

The project's starting point was a private LLM running inside an isolated macOS user account, hardened so it could not be modified. The intake role had been comparing the room's contents against a reference model on each visit; this installment turns that role into a vault keeper with a fixed checklist. Two concrete gaps surfaced along the way. The Mac's search feature showed an unhelpful "unknown" status, so the developer skipped the display and checked the index settings and index location directly. Each time the vault was closed, the intake and the vault files returned to the automatic backup target, and the settings were pinned so the room's inhabitants cannot undo them. A further snag showed that a watchdog test never really stopped the AI: the AI program exited on its own when the connection dropped, so the watchdog arrived to find nothing to kill. The developer rebuilt the decoy in Python, then found that the keeper and watchdog's permissions block markers on Apple-made programs, a platform behavior the post flags as a remaining item. Next time, the plan is to build another enclosure inside the room so that even tools the AI launches can be watched together.