Dynamic Self Stabalizing Loops (github.com)

miniBrain is a compact simulation lab for exploring recurrent neural-like dynamics focused on bistability, hierarchical reflection, and self-referential workspace models. It ships three related architectures: (A) bistable units coupled through a global workspace, (B) a reflective hierarchical “why-loop” and (C) a workspace that builds and predicts a compressed self-model. The environment is designed to probe stability, complexity and autocalibration: a background meta-tuner (optional PyTorch NN; falls back to heuristics) biases parameters toward high complexity (entropy + phase coherence), while irrational-time perturbations with drift prevent resonant locking. Built-in instrumentation computes Shannon entropy, a Lyapunov proxy, Lempel–Ziv complexity and pairwise mutual information to track dynamics quantitatively. Practically, miniBrain provides an interactive 3x2 GUI showing real-time heatmaps and R-phase plots (rolling window default 500 steps), plus a headless mode that exports CSVs and supports quick smoke tests for reproducibility. It's Python 3.10+ and depends on NumPy, Matplotlib, scikit-learn, and optionally PyTorch. Run with python3 lab.py to see autotuning in action, or run the provided smoke_test_autotune for lightweight validation. For researchers studying workspace theories, self-modeling, meta-learning or dynamical systems in ML, miniBrain offers a reproducible, extensible testbed with editable parameters (n_layers, dt, ROLLOUT_STEPS, reward weights) and documented findings in docs/findings.tex.