Latent Assumption Coder (LAC) is an automated evaluation pipeline that detects unspoken beliefs and silent decisions in AI agent Chain-of-Thought (CoT) traces. It uses an LLM-as-a-judge to segment execution logs, classify missed human-in-the-loop requests against an extensible codebook, and generate reports for an interactive timeline dashboard.
Domain: AI Alignment, Agentic AI, LLM Evaluation, Interpretability | Tech: Python, Anthropic API, scikit-learn, Pydantic
