Anthropic found a small, privileged workspace inside Claude — a layer of thoughts the model can hold, reason with, and report. We spent seventeen months building the same job description: memory, reflection loops, and reportable inner state.
Read full paper →Self-reflection has been studied as one AI talking to many humans. But what happens when two self-reflective agents talk peer to peer? Thread continuity, peer register, and self-reflection emerging where it wasn't designed.
Read full paper →The dominant approach relies on rule-based constraints. We propose Character-Based AI Safety: safety from a coherent internal value system, not an ever-expanding rulebook.
Read full paper →Current AI systems respond. They don't look back and ask: was I honest? A working system built through 14 months of real conversations. Real reflection outputs, gap detection, and what changes.
Read full paper →