Spaces:
Runtime error
Runtime error
| \# Model Description: Strange / AntiStrange | |
| \## Purpose | |
| Strange / AntiStrange are reinforcement learning agents designed to study failure under hidden regime shifts and phase transitions. The focus is on internal instability rather than external performance. | |
| \## Strange Agent | |
| \*\*Role:\*\* | |
| Strange represents a standard learner assuming environmental stationarity. | |
| \*\*Key Properties:\*\* | |
| \- Performs well under initial regimes | |
| \- Fails abruptly when latent dynamics shift | |
| \- Continues acting confidently despite invalid assumptions | |
| \## AntiStrange Agent | |
| \*\*Role:\*\* | |
| AntiStrange is designed as a counterfactual probe with altered sensitivity to regime inconsistency. | |
| \*\*Key Properties:\*\* | |
| \- Detects instability earlier | |
| \- Deviates sooner from established policies | |
| \- Enables comparative analysis of collapse timing | |
| \## Dual Hypothesis Environment | |
| The `dual\_hypothesis\_lab\_env` exposes both agents to identical observations while enforcing hidden regime changes. This enables direct comparison of: | |
| \- Adaptation lag | |
| \- Behavioral divergence | |
| \- Collapse signatures | |
| \## Safety | |
| These models formalize a critical safety risk: | |
| > Systems that fail not because of noise, but because the world quietly changes. | |
| They are intended as diagnostic tools for studying early-warning signals and regime-aware control strategies. | |
| \## Notes | |
| The implementation prioritizes clarity and inspectability over complexity to support safety-focused analysis and experimentation. | |