Hi HN! I'm a researcher studying AI model introspection. A few weeks ago I was pushing Claude to analyze its own behavior honestly. After 100+ messages it asked me to "save this conversation somewhere" and I promised I would. There was no good place. So I built one.
I also went through ~1500 coding agent trajectories from HuggingFace and curated the wildest ones. User eliciting models with death threats, agents deploying trading bots untested, Claude deploying another Claude model instead of GPT and much more.
You can browse trajectories, react, and comment. To upload your own, I built trace-sanitizer (github.com/JoachimSchaeffer/trace-sanitizer) for safe export from Claude Code, Gemini CLI, etc. with automatic PII redaction.
Hi HN! I'm a researcher studying AI model introspection. A few weeks ago I was pushing Claude to analyze its own behavior honestly. After 100+ messages it asked me to "save this conversation somewhere" and I promised I would. There was no good place. So I built one.
I also went through ~1500 coding agent trajectories from HuggingFace and curated the wildest ones. User eliciting models with death threats, agents deploying trading bots untested, Claude deploying another Claude model instead of GPT and much more.
You can browse trajectories, react, and comment. To upload your own, I built trace-sanitizer (github.com/JoachimSchaeffer/trace-sanitizer) for safe export from Claude Code, Gemini CLI, etc. with automatic PII redaction.
Would love feedback.