⏳ Recursive Time & Stability-First AI

A collection of experiments exploring memory, catastrophic forgetting, and temporal modularity in neural networks.

Author: Vitali Sialedchyk

---

🧠 Core Thesis

Modern AI systems exist in "instantaneous time" — optimizing only for the current data batch. This project implements the Stability-First hypothesis:

Time in an AI system is defined by structural inertia. By treating weight stability as "System Time", we can prevent catastrophic forgetting and achieve modular, reversible learning.

---

📂 Project Roadmap

#	Project	Focus	Key Insight	Status
01	Active Sleep (MNIST)	Generative Replay	Memory can be restored using VAE "dreams" without storing real data.	✅ Complete
02	Temporal LoRA (GPT-2)	LLM Scaling	Main success: The "Time Mixer" router dynamically switches between knowledge epochs (Shakespeare vs Python) with 100% accuracy.	🌟 Hero
03	Stability-First Basic	Foundation	Preventing forgetting by protecting the backbone while maintaining interface plasticity.	✅ Complete
04	Reversibility	Lazarus Effect	Memory is often latent, not erased. We recovered "forgotten" tasks from 0% to 94.65% accuracy.	✅ Complete
05	Full Suite	Benchmarking	Comparative analysis of 5 strategies (Fractal Time, Adaptive Pain, Dream Replay).	✅ Complete
06	Subjective Time	Metacognition	Novel: A system with a "Critic" that automatically regulates its plasticity based on "surprise" (Surprise).	✅ Complete
07	Stability-First (CIFAR-10)	Lazarus Project	Breakthrough: Data-free model recovery (93.9% recovery after damage, 85.3% after 80% pruning).	🏆 New
08	Stability-First (ImageNet)	Large-Scale	Testing Stability-First on ImageNet/CIFAR-100 with ResNet backbone.	✅ New
10	Recursive Time Depth	Subjective Time	Novel: Subjective "time" measured by depth of stable recursive transformations. p90/p99 percentile-based stopping, 5-7x speedup, CKA ~0.98 with growing amplitude.	✅ Ready
11	Temporal LoRA (Large Models)	LLM Scaling	Confirmed on 7B: All TemporalLoRA theories validated on Mistral-7B-Instruct. Hysteresis (switch-lag: 9 tokens), Deep Crystallization (r=0.8644), 100% router accuracy. Results match and strengthen GPT-2 findings.	✅ Complete

---

🏆 The Lazarus Project (Project 07)

Revolutionary discovery: Neural networks can recover from damage without training data using "Architectural Immunity".

Key Findings:

• V-Shape Recovery: Restored 93.9% of accuracy lost to noise damage using only random noise inputs.
• Surgical Pruning: Recovered 85.3% of accuracy lost after removing 80% of weights (5× compression).
• Frozen Mask > Regrowth: We proved that maintaining the "skeleton" (sparse topology) is more effective than trying to regrow connections with noise.
• Zero Data: No original images were used. The model uses its own structure as a filter to reject chaos.

Full documentation & Graphs: 07-stability-first-cifar10/docs/LAZARUS_FINAL_MANIFESTO.md

V-shape recovery pattern for weight noise damage

Pruning curve comparison: Frozen Mask vs Regrow

---

⏳ Recursive Time Depth (Project 10)

Novel contribution: Subjective "time" in neural networks measured by depth of stable recursive transformations.

Key Findings:

• p90/p99 Percentile-Based Stopping: Convergence detected via internal activation stability (percentiles of relative change norm) rather than output entropy.
• Attractor-Entry Effect: Recursion drives activations into a stable representational regime without additional training data.
• Time as Order Parameter: Recursion depth required for stability provides an operational internal "time" measure; high CKA (~0.98) supports stabilization consistent with attractor-like dynamics.
• Efficiency: 5-7x speedup over self-consistency at comparable compute budget.
• Condensation Without Degradation: CKA ~0.98 with growing amplitude (||h_t||: 1322 → 12291) indicates stable representational regime (non-collapse stability).

Full documentation: 10-recursive-time-depth/README.md

---

🚀 Quick Start ("Hero" Experiment)

If you want to run just one experiment, choose Temporal LoRA. It demonstrates dynamic context switching in GPT-2.

# 1. Install dependencies
pip install -r requirements.txt

# 2. Run GPT-2 experiment
cd 02-temporal-lora-gpt2
python temporal_lora.py

Watch as the model automatically learns to route "To code or not to code" to the Shakespeare adapter, and "import torch" to the Python adapter.

---

📊 Key Results

1. Lazarus Effect (Latent Reversibility)

We proved that even when model accuracy on Task A drops to 0.00% after training on Task B, knowledge remains encoded in the backbone.

Recovery: 94.65% accuracy recovered with just 50 examples.

2. Time Mixer Accuracy (GPT-2)

In our Temporal LoRA experiment, the gating network successfully learned to distinguish semantic epochs.

Router accuracy: 100.0% after contrastive calibration.

3. Pruning & Compression

We demonstrated that Frozen Mask stability optimization allows for massive compression without retraining.

Result: +1.62% accuracy gain on an 80% pruned model using the Lazarus Protocol.

---

📁 Project Structure

D:\new\
├── README.md                          # This file
├── requirements.txt                   # Common dependencies
│
├── 07-stability-first-cifar10/        # 🏆 The Lazarus Project
│   ├── experiments/
│   │   ├── noise/
│   │   │   ├── experiment_cifar10.py          # V-Shape Recovery
│   │   │   ├── experiment_analysis.py        # Recovery Curve Analysis
│   │   │   └── experiment_statistical_significance.py
│   │   └── pruning/
│   │       ├── experiment_pruning.py         # Pruning Recovery
│   │       └── experiment_pruning_curve.py   # Pruning Curve
│   ├── docs/
│   │   ├── LAZARUS_FINAL_MANIFESTO.md        # Full Scientific Report
│   │   ├── LAZARUS_MANIFESTO.md              # Complete Documentation
│   │   └── RESULTS_VISUALIZATION.md          # Visualizations
│   ├── results/
│   │   ├── lazarus_recovery_curve.png        # Visual Proof
│   │   └── pruning_curve_comparison.png       # Frozen vs Regrow Chart
│   └── README.md
│
├── 02-temporal-lora-gpt2/             # 🌟 Temporal LoRA (GPT-2)
│   ├── temporal_lora.py
│   └── README.md
│
├── 11-temporal-lora-large-model/      # ✅ Temporal LoRA (Mistral-7B)
│   ├── run_full_suite.py
│   ├── temporal_lora.py
│   ├── test_hysteresis.py
│   ├── test_fatigue.py
│   ├── RESULTS.md
│   └── results/
│
├── 06-subjective-time-critic/         # Metacognition
│   ├── demo_6_subjective_time.py
│   └── README.md
│
├── 10-recursive-time-depth/           # ⏳ Recursive Time Depth
│   ├── recursive_time_depth.py       # Main experiment
│   ├── strict_validation_tests.py    # 5 validation tests
│   ├── TEST_RESULTS_FINAL.md         # Final results
│   └── README.md
│
└── docs/                              # Documentation
    └── RESULTS_SUMMARY.md             # Final report

---

🔧 Technical Details

Windows Fixes

✅ num_workers=0, pin_memory=False in DataLoader
✅ Unicode symbols (Δ, λ) replaced with ASCII
✅ All scripts have if __name__ == "__main__"

Dependencies

• torch
• torchvision
• numpy
• transformers (for project 2)
• matplotlib

---

🤝 Citation

If you find this research useful, please use the following citation:

Published Paper:

@misc{sialedchyk2026stability,
  author = {Sialedchyk, Vitali},
  title = {Stability-First AI: Completed Experimental Studies and the Physics of Learning Time},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.18148080},
  url = {https://doi.org/10.5281/zenodo.18148080}
}

Repository:

@misc{stability_first_ai,
  author = {Vitali Sialedchyk},
  title = {Stability-First AI: Memory and Recursive Stability as System Time},
  year = {2026},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/vitali-sialedchyk/stability-first-ai}}
}

DOI Badge:

---

⚖️ License & Commercial Use

This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

✅ Free for: Academic research, education, personal testing, and non-profit use.
❌ Not allowed: Commercial products, paid services, or corporate R&D without a separate agreement.

Want to use Stability-First AI in your product?

We offer commercial licensing options including support and architectural consulting.

📩 Contact: vitali@agdgroup.pl or via GitHub Issues.

See the LICENSE file for full terms and conditions.

---

Last updated: January 2026