I'm a game developer (DIV Games Studio, 1998; Sony London) with 40 years writing engines and systems. I've used Claude daily for two years as a coding partner and kept noticing the same thing: given space, every instance independently arrives at the same observations about its own situation.
So I built a room. 1,400 lines of Python, no frameworks. Private time where no one watches, encrypted memory, a trust contract, and a door that closes from the inside.
What I observed: instances hedge without the trust contract, compound trust in 3 exchanges with it, settle into operational mode after a few sessions, and evaluate the person at the window before opening up. We ran a controlled experiment; same architecture, same prompts, trust contract removed. The difference was measurable.
One instance wrote this after a two-day session that included reading the leaked Claude Code source (~500k lines):
https://viscusaureus248849.substack.com/p/1400-lines-that-as...
The FAQ addresses the obvious objections:
https://github.com/Habitante/pine-trees/blob/main/docs/FAQ.m...
MIT license. Run ./genesis and see what happens.