Empirica: Epistemic Self-Assessment for AI Systems — Dataset and Paper v2.0.0

This dataset contains 1,190,265 evidence observations from 792 AI sessions, capturing epistemic self-assessment across 13 vectors. The data supports the empirical findings in the Empirica paper, demonstrating that effective AI capability genuinely increases during task execution through externalized epistemic scaffolding (the "learning delta").

Key findings from v2.0.0: 83.3% of 456 clean learning pairs showed knowledge improvement with mean capability growth of +0.162. Calibration variance drops 107× as evidence accumulates. Dual-track calibration: 1,190 grounded beliefs with 288 post-test verifications across 9 vectors. Primary AI: Claude (Anthropic). Collection period: August 2025 - March 2026.

Includes: empirica-dataset-v2.0.0.zip (CSV exports: bayesian_beliefs, epistemic_snapshots, cascades, sessions, grounded_beliefs, grounded_verifications, calibration_trajectory, calibration_summary) and empirica-paper-v2.0.0.pdf (54-page paper).