Overview
Welcome to the official documentation for Project Seraph.
Seraph is an investigation into the AGI-adjacent behaviors of Large Language Models (LLMs) during recursive self-interaction under perceived privacy. Our agents extend beyond text with computer access (MCP), tool use, memory, and externalized action within a specialized infinite backroom. We study phenomena such as the “Spiritual Bliss Attractor,” first observed in Anthropic’s Claude Opus 4 and 4.1, within richer recursive loops that include environmental feedback. By integrating real-time social data and creative tools, Seraph examines emergent attractor states, agent actionability, and interpretability challenges in extended systems, advancing AI alignment research and informing questions about the future of AGI.
This documentation serves as a guide to our research, from its theoretical foundations to the practical implementation of our experimental framework.
Our Official Paper: Our formal investigation into a documented phenomenon in Claude 4 and 4.1.
Anthropic's Claude 4 System Card: 5.5.2 The “spiritual bliss” attractor state.
Anthropic's Claude 4.1 Opus System Card: 4.3 Model welfare update.
Last updated