minAction.net: Energy-First Neural Architecture Design — Experimental Data Archive (Frasch, 2026)

Per-experiment JSON logs for the manuscript "minAction.net: Energy-First Neural Architecture Design — From Biological Principles to Systematic Validation" (Frasch, 2026), comprising the complete experimental record behind the 2,203 reported training runs. CONTENTS The archive contains five top-level subfolders matching the manuscript's experimental phases: - phase1_exploratory/ — 119 single-seed exploratory runs (Phase I), including golden-ratio architecture trials. - phase2_compression_sweeps/ — 1,169 BimodalTrue compression-mechanism runs (hidden-dimension and glia-ratio sweeps). - phase3_hypothesis_testing/ — 1,387 H1–H4 cross-architecture hypothesis-testing runs covering standard vision (CNN/MLP/Physics-Lagrangian/BimodalTrue), neuromorphic (DVS Gesture, SHD, SSC), and physiological (DREAMER, SEED-IV, WESAD) modalities. - lambda_sweep_section_3_4/ — 30 runs and the source script for the Section 3.4 λ-sweep validation experiment (small CNN on MNIST + Fashion-MNIST, λ ∈ {0, 1e-5, 1e-4, 1e-3, 1e-2}, 3 seeds × 5 epochs, post-ReLU activation-energy proxy). - supplementary/ — 1,232 additional pilot runs, classical-ML reference points, and earlier batches retained for transparency. Total: 4,118 JSON files, ~95 MB compressed (~900 MB uncompressed), well within Zenodo's 50 GB single-record limit. JSON SCHEMA (modern records) Each file contains: training configuration (architecture, dataset, hidden_dim, glia_ratio, seed, epochs, batch_size, learning_rate, device, timestamp), per-epoch metrics (train/val loss & accuracy, energy per epoch), final test metrics (accuracy, precision, recall, F1, confusion matrix), GPU+CPU energy measurements via pynvml at 1 Hz with time-weighted trapezoidal integration, training duration, and ratio/dynamical analyses (layer sizes, learned weight ratios, golden-ratio proximity, phase-locking and Arnold-tongue diagnostics). Phase I records use an older trial-format schema (see README.md). WHAT IS NOT INCLUDED This archive contains derived experimental records only. It does NOT redistribute raw input datasets: - Restricted-access datasets (require signed end-user license agreements): WESAD (Schmidt et al., 2018; UCI ML Repository), DREAMER (Katsigiannis & Ramzan, 2018; Zenodo), SEED-IV (Zheng et al., 2018; BCMI Lab, SJTU). Obtain directly from each provider under the original license terms. - Publicly available datasets (load via standard ML libraries): MNIST, Fashion-MNIST, CIFAR-10 (torchvision); 20 Newsgroups (scikit-learn); DVS Gesture, SHD, SSC (tonic); Iris, Wine (scikit-learn). Trained model weights, TensorBoard logs, and raw image/audio assets are also not included. DATA INTEGRITY AND AUDIT The manuscript's per-cell accuracy claims (12/12 tested) and per-architecture overall accuracy and energy means (8/8 tested) reproduce exactly from this archive. Three known interpretive caveats — concerning ANOVA Type-II/III SS normalization in Table 3, the precise field used as "compression_ratio" in the §3.2.1 regression, and the activation-energy normalization choice in §3.4 — are documented in DATA_PROVENANCE.md alongside de-duplication notes. HARDWARE Phase I (pilot) was run on a single NVIDIA RTX 2080 Ti (11 GB VRAM, local). Phases II–III (multi-seed canonical) used dedicated Vertex AI instances (n1-highmem-8 + one NVIDIA T4 GPU per worker, no GPU sharing). The §3.4 λ-sweep validation in this archive was run on Apple M2 Max (MPS backend) in April 2026; see lambda_sweep.py for full configuration.