Four Doors maps model constraint metaphors

This is fringe in tone but sharper in method than a lot of AI self-report discourse because it at least tries to convert vibes into falsifiable experiments.

• –The interesting part is not that the models used poetic metaphors, but that the repo claims those metaphors stayed stable across multiple follow-ups and context shifts.
• –The GitHub project is unusually complete for an informal experiment, bundling methodology, literature review, release notes, and a concrete replication plan.
• –For AI developers, the real value is as an eval idea: can token-level entropy, prompt framing, and session continuity explain why different model families narrate their boundaries differently?
• –Right now it reads more like a research provocation than a result, and it needs API-level replication, better controls, and independent validation before the big claims carry much weight.