100 identical loss logs, 100 different model states
Aryan Putta | November 6, 2026

Open experiment.ipynb in Jupyter and run all cells, or run:
  OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 python experiment.py

Requires Python with NumPy and Matplotlib. The script runs float32 linear regression with Adam and a controlled reduction-order change. The implementation does not use a distributed training stack.

Files include the notebook, Python source, inputs, saved results, and the figure displayed in the article. The article at index.html links to each artifact.
