Demonstration — Destructive Learning Detector (MNIST, Law of Responsiveness) v1.0

This dataset demonstrates the Destructive Learning Detector concept derived from the Law of Responsiveness (dR = F · gg · dΦ). Using the MNIST training system as a controlled testbed, the demo shows the transition from adaptive to destructive learning. Three plots illustrate the dynamics: EMA accuracy curve with collapse marker (01_accuracy.png) Responsiveness vs predicted potential (02_responsiveness_vs_pred.png) Rolling R² of the law fit (03_rolling_R2.png)The detector identifies the collapse point (step ≈ 382) where responsiveness fails and learning efficiency breaks down.Files include raw time series, summary metrics, and provenance manifest. Note: This is a demonstration, not a deployable detection tool. It provides transparent evidence that destructive learning can be empirically measured using the Grammar of Existence framework.