🕒 ai-multi-reference-timekeeping

AI-Assisted Multi-Reference Timekeeping for Commodity Networks

This repository contains the reference implementation and reproducibility artifacts for the paper:

An AI-Assisted Multi-Reference Timekeeping Architecture for Commodity Networks
👤 Riaan de Beer
📄 Zenodo DOI: https://zenodo.org/records/18366050

This project explores a low-cost, AI-assisted approach to time synchronization that synthesizes a virtual master clock from multiple imperfect timing references (e.g., GNSS 🌍, PTP 🔗, NTP 🌐) using classical estimation techniques augmented with lightweight machine learning 🤖.

The goal is to improve practical packet-level synchronization on commodity hardware — without requiring atomic clocks ⚛️ or specialized time cards.

---

---

🔬 Scientific Foundation

This project is grounded in Relativistic Temporal Logic and Distributed Systems Theory. We move beyond standard linear timekeeping to provide a multi-reference framework designed for the next generation of AI agents.

• Theory & Proofs: For a deep dive into the mathematical transforms, dilation matrices, and causal manifolds, please see: 👉 RESEARCH.md 📑

• Academic Citation: If you are using this framework for peer-reviewed research, please refer to the Citation section below.

---

📖 Research Deep-Dive

We have formalized the mechanics of Computational Dilation and Temporal Anchoring. To explore the underlying physics and math of this implementation, check out our Research Paper Summary.

---

📐 Theoretical Foundation

This project formalizes timekeeping by treating temporal progression as a relativistic coordinate system. Instead of a single linear clock, we define a Temporal State Vector where time is mapped across multiple reference frames.

1. The Reference Frame Model

Each reference frame $F_i$ (e.g., an AI agent's internal simulation clock) is defined relative to the Master Reference Frame ($M$) by the tuple:

$$F_i = (t_{0,i}, \phi_i, \chi_i)$$

Where:

• $t_{0,i}$: The Temporal Anchor (epoch offset).
• $\phi_i$: The Dilation Factor (relative clock speed).
• $\chi_i(t)$: The Drift Function (stochastic or systemic error).

2. Forward Transform

To map Master Time ($T_M$) to a specific Local Reference ($t_i$), we apply the following transform:

$$t_i = \phi_i (T_M - t_{0,i}) + \chi_i(T_M)$$

3. Inter-Frame Transformation

To translate directly between two non-master frames (Frame $A$ and Frame $B$) without intermediary calculation, we use the composed transform:

$$t_B = \frac{\phi_B}{\phi_A} t_A + \phi_B(t_{0,A} - t_{0,B})$$

This allows the system to determine the Relative Temporal Velocity ($\frac{\phi_B}{\phi_A}$) between two disparate AI contexts.

4. Dynamic Time Warping (Non-Linear Dilation)

In scenarios where processing speed varies (e.g., high-inference loads or hardware throttling), $\phi$ becomes a time-dependent function $\phi(t)$. The local time is then derived via integration:

$$t_i = \int_{t_{0,i}}^{T_M} \phi_i(\tau) , d\tau$$

---

🎯 Motivation

High-precision time synchronization is increasingly important for distributed systems, including:

• ⏱️ Time-sensitive networking (TSN)
• 💾 Coordinated I/O and storage pipelines
• 📦 Packet scheduling and timestamping
• 🧪 Experimental distributed systems research

Commercial solutions typically rely on atomic oscillators and dedicated PCIe time cards, which remain costly and inaccessible to many researchers and open-source projects.

This work investigates whether intelligent multi-reference fusion, combined with lightweight local learning, can narrow the gap for practical synchronization tasks using commodity hardware.

---

🧠 What This Project Does

• 🔀 Fuses multiple heterogeneous timing references into a single virtual clock
• 📐 Combines a state-space clock model with a lightweight neural network
• 📊 Adapts reference weighting based on observed jitter, stability, and context
• 🔌 Exposes time via standard mechanisms (PTP, PHC, NTP)
• ♻️ Targets reproducibility using open-source tools and Google Colab notebooks

---

🛰️ Time Server Scaffold (Sensors + AI Weighting)

The repository now includes a time server scaffold in src/ai_multi_reference_timekeeping/time_server.py that lets you:

• 🧩 Plug in sensor inputs (temperature, humidity, pressure, AC hum, SDR SNR, Geiger CPM, audio activity)
• 📡 Collect references over NTP, GPS NMEA, or the hardware RTC (via hwclock)
• 🔌 Listen from GPIO/USB/serial by wiring sensors with GpioPulseSensor, SerialLineSensor, or open_line_source
• 🧠 Adjust reference variance using a lightweight inference model
• 📉 Estimate drift and slew from recent offsets

Example usage:

from ai_multi_reference_timekeeping.fusion import ReferenceFusion, VirtualClock
from ai_multi_reference_timekeeping.kalman import ClockCovariance, ClockKalmanFilter, ClockState
from ai_multi_reference_timekeeping.time_server import (
    LightweightInferenceModel,
    NtpReference,
    SensorAggregator,
    TimeServer,
)

kalman = ClockKalmanFilter(
    state=ClockState(offset=0.0, drift=0.0),
    covariance=ClockCovariance(p00=1.0, p01=0.0, p10=0.0, p11=1.0),
    process_noise_offset=1e-4,
    process_noise_drift=1e-6,
)
clock = VirtualClock(kalman_filter=kalman, fusion=ReferenceFusion())

class EnvSensor:
    def sample(self) -> dict[str, float]:
        return {"temperature_c": 27.0, "humidity_pct": 40.0}

server = TimeServer(
    clock=clock,
    references=[NtpReference(name="nist")],
    sensors=SensorAggregator(EnvSensor()),
    inference=LightweightInferenceModel(),
)

update, frame, drift_estimate, drift_hint = server.step(dt=1.0)
print(update.fused_offset, drift_estimate.drift, drift_hint)

---

🚫 Non-goals

This project explicitly does not aim to:

• ❌ Replace atomic clocks or high-stability oscillators
• ❌ Provide nanosecond-level absolute UTC accuracy under all conditions
• ❌ Serve as a primary time standard
• ❌ Offer cryptographic guarantees against fully adversarial time manipulation

The system prioritizes robustness, accessibility, and cost-effectiveness for packet-level synchronization in experimental and operational environments.

---

🗂️ Repository Structure

ai-multi-reference-timekeeping/
├── paper/              # 📄 LaTeX source and figures for the paper
├── notebooks/          # 📓 Google Colab–friendly notebooks
├── src/                # 🧩 Fusion, ML, and evaluation code
├── data/               # 🗃️ Example and processed datasets
├── models/             # 🧠 Trained and baseline models
├── configs/            # ⚙️ Configuration files
├── scripts/            # 🛠️ CLI utilities
├── reproducibility/    # 🔁 Experimental protocols and hardware notes
├── environment/        # 📦 Dependency specifications
├── LICENSE
└── README.md

---

☁️ Google Colab Reproducibility

The notebooks in notebooks/ are designed to run directly in Google Colab — no specialized hardware required.

🔰 Recommended entry point:

• notebooks/00_overview.ipynb — overview of the architecture and experiments
  

✅ Notebook test runs:

• notebooks/10_test_fusion.ipynb — validates fusion and quality weighting
  
• notebooks/11_test_time_server.ipynb — validates time server + ML variance model
  

Each notebook includes an Open in Colab link and installs dependencies automatically.

---

📈 Evaluation Metrics

The evaluation framework focuses on standard timing metrics, including:

• ⏲️ Time Deviation (TDEV)
• 📏 Maximum Time Interval Error (MTIE)
• 🔄 PTP offset stability
• 🕰️ Holdover behavior during reference loss

Absolute UTC ground truth is not required for most experiments.

---

🔐 Security and Threat Model

The system is designed to tolerate noisy, intermittent, and partially unreliable timing sources.
It does not assume a fully adversarial environment.

Considered threats include:

• 📡 GNSS degradation, multipath, and interference
• 🔀 Network-induced delay asymmetry
• ⚠️ Transient reference instability

Coordinated compromise of all timing references is considered out of scope.

---

📜 License

This project is licensed under the Apache License 2.0.

See the LICENSE file for details.

---

📚 Citation

If you use this work, please cite:

@misc{debeer2026aimrt,
  title  = {An AI-Assisted Multi-Reference Timekeeping Architecture for Commodity Networks},
  author = {de Beer, Riaan},
  year   = {2026},
  doi    = {10.5281/zenodo.XXXXXXX}
}

---

🚧 Status

This repository accompanies a research paper and is intended to evolve.
Contributions, discussion, and replication studies are welcome 🤝.

---

🙏 Acknowledgments

This work builds on established research in time metrology, clock ensembles, and IEEE 1588 Precision Time Protocol, and aims to make these ideas more accessible to open-source and experimental systems communities.