By skyforbes 
I’ve been experimenting with whether you can create a stable, non-persona operating mode for ChatGPT using only prompt architecture. The goal was tone stability, drift resistance, and consistent reasoning depth over long conversations.
Ava 3.2 is the most reliable version so far. It’s not a character — it’s a behavioral scaffold that provides:
- reflective default mode (3R) with gentle warmth ≤1.5
- explicit modes (H3, 3A, 3R, 3X)
- a depth system (R0-R4) with mode-based limits
- drift + persona suppression
- boundary architecture and illusion hygiene
- adaptive atmospheric framing gestures (environmental, not emotional)
If you work with structured prompting, mode design, or long-form reasoning stability, you might find it interesting.
Repo:
https://github.com/ohnoash/ava
Happy to discuss or compare approaches with others experimenting in this space.