Structural Time in Machine Learning: A Two-Part Preprint Series on Criticality-Driven Compression and LoRA Scaling (with Supplementary Code & Data)

This unified deposit contains the complete research work, including both preprints and the fully curated supplementary code and experimental data, for the Structural Time Framework in Machine Learning (STF-ML) project. 1. Paper 1: Core Framework "Structural Time in Machine Learning: Replacing Smooth Gradient Descent with Criticality-Driven Compression" Proposes mapping the latent representations of neural networks to a nonlinear dynamical system where temporal density $T(K)$ acts as a metric for compression efficiency. Validated on character-level transformers (~421K parameters) with Adaptive Criticality Injection showing consistent improvements ($t(4) = 4.89, p = 0.004, d = 2.19$). 2. Paper 2: LoRA Scaling "Structural Time in Machine Learning: Scaling Criticality-Driven Compression via LoRA Fine-Tuning" Scales the STF methodology to production-scale models (GPT-2 124M, TinyLlama 1.1B, BERT-base 110M, Qwen3.5 0.8B) using LoRA. Demonstrates that Continuous Hybrid Steering prevents representation collapse (OOD Perplexity reduced by 92.9% vs. Control), introduces the Triad Evaluation Pipeline, and identifies the Structural Camouflage Paradox (the allostatic toll of sustained Critical-regime training). 3. Supplementary Code & Data Folder Includes all Google Colab-compatible training scripts, model configs, raw benchmark logs, specs (THM probe and Triad specifications), and the main figures reporting the results. Deposit Files: structural-time-in-ml-core.pdf (Paper 1) structural-time-in-ml-lora-scaling.pdf (Paper 2) supplementary.zip (Curated code, results, and specs)