KiRoJota: A Hierarchical Family–Template–Instance Architecture for Physical Law Rediscovery from Sensor Data

We present KiRoJota, a proof-of-concept system for discovering, validating, and reusing physical laws from data. Its central architectural contribution is an explicit Family → Template → Instance hierarchy that separates the mathematical structure of a physical law from the experiment-specific parameters fitted to individual observations. The primary evaluation is an end-to-end experiment on simulated LiDAR and Doppler sensor data for free-fall motion, including a causality benchmark (200 scenes, g reconstructed to within 0.03% of the true value) and an anomaly benchmark (50 scenes with air resistance, 0 cases of silently reusing an invalid law). A second, intentionally narrower experiment shows the same architectural principle transfers across four physical domains (mechanics, electronics, biology, thermodynamics; 14 templates) with 97.5% accuracy on mathematically distinguishable cases. The paper reports both successful results and observed limitations, including mathematically non-identifiable model families, noise sensitivity, and every non-trivial implementation defect encountered during development.