About
About Paul Edward Harris 4th
I'm an independent AI/ML learner and researcher from Connecticut, studying on Cape Cod, Massachusetts. I don't have a lab or a CS degree yet — what I have is a habit of building things that are a little past what I currently understand, then working until I do understand them.
My process looks like this:
- idea
- research
- prompting AI tools to help generate code
- running it and hitting errors
- investigating and correcting
- testing
- documenting what actually happened
Right now my focus is learning my own projects from the inside out — understanding every file well enough to explain it and change it without AI writing the change for me.
Research
Chaotic forecasting under irregular sampling
Continuous-Time Latent Dynamics for Long-Horizon Chaotic Forecasting under Irregular Sampling. Paul E. Harris IV, Independent Researcher, Mashantucket Pequot Nation. SynechismCore Technical Report v1.1.0, November 2026 (15 pages). DOI: 10.5281/zenodo.23772387.
Architecture. SynechismV23 is a forecasting architecture I designed. It is a latent neural ODE: instead of updating on a fixed step, it learns a continuous rule for how its hidden state changes over time, so it can forecast to any point in time. It adds an elastic attractor term that keeps long forecasts bounded, golden-ratio sub-step scheduling for the integrator, and a laminar bypass that switches to a simpler linear update where the dynamics are smooth.
Question. When observations arrive at uneven times, does SynechismV23 keep chaotic forecasts accurate longer than three sequence-model baselines (a Transformer, an LSTM and a simplified Mamba-style state-space model) given the same timing information?
Setup. The Lorenz-63 system observed under four sampling regimes with the same nominal average rate: regular, jitter, half missing and exponential. Every model was trained for 200 epochs, over 5 seeds, and scored by valid prediction time (VPT) in Lyapunov times.
Results. SynechismV23 had the longest valid prediction time in all four regimes. The Transformer and Mamba-style baselines lost about 80% of their VPT between regular and exponential sampling. Against the LSTM, SynechismV23's lead under irregular sampling is small (4 of 5 seeds, 9–14%) and not yet statistically significant across seeds; more seeds are needed. The LSTM degrades least as timing becomes uneven; SynechismV23 ranks highest because it starts much higher.
VPT under exponential sampling (Lyapunov times)
| Model | VPT |
|---|
| SynechismV23 | 2.03 |
| LSTM | 1.78 |
| Transformer | 0.55 |
| Mamba-style | 0.50 |
Limits. One system (Lorenz-63); a separately trained model per sampling regime; the Transformer and Mamba-style baselines output a whole forecast block while SynechismV23 and the LSTM roll forward, so architecture and decoding are not yet separated; the Mamba-style model is a simplified implementation; component ablations are so far at 50 epochs only; the report has not been peer reviewed.
Next steps. Matched step-by-step baselines, a reference Mamba and a continuous-time Transformer, more seeds, full-budget ablations, a turbulence benchmark (Kuramoto–Sivashinsky), real robotics data, then arXiv and a 2027 workshop submission.
Project history. SynechismCore began as an exploratory study of forecasting under periodic correction. That study was discontinued, and this report supersedes it; only the irregular-sampling results above are claims of this work.
Paper (PDF) · Zenodo DOI · Code and per-forecast results (MIT)
How this was made: this site and my projects are built with heavy AI assistance. I direct the work, run it, debug it, and check it — and I'm actively learning to understand every part myself.
