Start with the character, not the machinery. Talk to a published persona and watch governed memory change over time. Then flip the conversation to inspect the exact versioned file that produces the voice.