What it is
An agentic interpretability system that autonomously generates, runs, verifies, and iterates on mechanism-discovery experiments for AI models. Unlike AI-scientist systems aimed at task solutions, it targets theory-level mechanism discovery: grounded by a 13k-paper interpretability KG, 43M-paper corpus, and 32 analysis tools, it proposes, tests, verifies, and iterates on mechanistic hypotheses.
Key results
- Belief-head intervention improves Pythia-410M by +15.3% over baseline
- Unsafe-response rate hits 48.6% vs 20.3% baseline, 18.3% control
- Banana generation rises to 25.6% vs 2.5% baseline, 2.1% control
Mechanistic hypotheses can be autonomously turned into reliable experiments that expose hidden risks and yield actionable control knobs.