Empirica Dataset: Epistemic Self-Assessment for AI Systems

This dataset contains 87,871 evidence observations from 852 AI sessions, capturing epistemic self-assessment across 13 vectors. The data supports the empirical findings in the Empirica paper, demonstrating that AI capability genuinely increases during task execution (the "learning delta"). Key findings: 91.3% of sessions showed knowledge improvement with mean KNOW delta of +0.172. Calibration variance drops 62× as evidence accumulates, demonstrating genuine Bayesian convergence. Primary AI: Claude (Anthropic). Collection period: August 2025 - January 2026.