Empirical Results — Test of the Laws of Persistence and Coherence (MNIST EWC) v1.0 - Result: conservation confirmed within empirical benchmark (ΔΣ ≈ 0.017 < threshold 0.18)

This dataset and analysis validate the Law of Persistence & Coherence proposed in Grammar of Existence — Part 2. Using the public MNIST continual-learning benchmark, we tested whether a learning system preserves structural balance when sequentially trained on related tasks. A small convolutional model (SmallCNN) was trained on digit pairs with and without Elastic Weight Consolidation (EWC). Diagonal accuracies (≈ 0.97–0.999) and mean |Σbalance| = 0.017 (< 0.18 threshold) confirm the predicted conservation relation dReg* + dRig* + dDist* ≈ 0 and dCoup* + dAlign* + dProp* ≈ 0. No falsifier was triggered. The EWC variant maintained coupling and memory stability, demonstrating persistence and coherence across learning transitions. Results reproduce the cross-domain balance pattern seen in prior synthetic runs, showing that the conservation law generalizes to empirical data. All scripts, plots, and metrics are included for full reproducibility.

------------------------------------------------------------------------------------------------This dataset accompanies the preregistration Pre-Registration: Laws of Persistence & Coherence (Grammar of Existence — Part 2) v1.0.

It contains MNIST split-5 continual-learning runs (baseline and Elastic Weight Consolidation [EWC]) used to illustrate the Persistence / Coherence diagnostic.

Files: accuracy matrices (CSV), plots (PNG), optional model weights (PT), and the Colab notebook.

Result: Mean |Σbalance| = 0.017 (threshold 0.18) → no falsifier triggered; diagonal accuracies ≈ 0.97–0.998 across stages, indicating persistence under EWC.

Reproduction: open the included notebook and run cells in order (PyTorch CPU OK).

Provenance: public MNIST dataset (torchvision).