\# Phase Shift \& Regime Uncertainty  

\## Strange / AntiStrange



This project studies reinforcement learning systems operating under hidden regime shifts and non-stationary dynamics. It focuses on how agents fail when the environment changes phase without explicit signaling, a scenario common in real-world deployment and safety-critical systems.



\## Core Idea



Many environments appear stationary until they are not. This project explores:

\- Latent phase transitions

\- Sudden behavioral collapse despite stable metrics

\- The gap between observed performance and underlying stability



Strange / AntiStrange are paired agents designed to probe these dynamics from different behavioral assumptions.





\## Project Structure



\- `agents/`  

  #Strange and AntiStrange agent implementations.



\- `envs/`  

  #Phase-shifting environments, including `dual\_hypothesis\_lab\_env`.



\- `run\_all.py`  

  #Executes both agents in a shared environment for direct comparison.



\- `results/`  

  #CSV logs capturing regime transitions and performance degradation.



\- `graphics/`  

  #Comparative plots showing divergence under regime shifts.





\## Running the Experiments



This project is designed to run without notebooks. (I resorted to Anaconda)



```bash

conda activate your\_env

python run\_all.py



\## Safety 

\## Why This Matters for AI Safety



-Hidden regime shifts are a major source of real-world AI failures. 

-This project operationalizes questions such as:



  -How long can agents operate under false assumptions?



  -What does “pre-collapse” behavior look like?



  -Can instability be detected before performance drops?



\## Related Work



-This project accompanies the preprint:

-Adaptive Stability Control in Sequential Learning Systems