---
pretty_name: SNO Article I Evidence
license: cc-by-4.0
language:
- en
tags:
- machine-learning
- reproducibility
- causal-analysis
- mnist
---

# SNO Article I Evidence

Machine-readable experimental records supporting *Watching Training Move:
Causal Forecasting of Neural Updates Across 8,192 SGD Transitions*.

Creator: **TiGa-RCE**  
Affiliation: Independent researcher; founder and sole operator of Brewster Jennings

The Dataset contains original result records, figure-data contracts, and
machine-readable receipts. It does not contain the article prose, the rendered
figures, the source code, private conversations, local system paths, model
caches, or downloaded MNIST data.

The long-horizon evidence includes a matched-runtime reproduction receipt for
the exact PyTorch 2.13 result through 8,192 transitions and separately records
the observed PyTorch 2.8 strict-hash portability boundary.

`evidence/horizon/zerogpu-portability/` adds a bounded three-horizon control
comparing the local PyTorch 2.13 CPU path with a PyTorch 2.13 CUDA run on
Hugging Face ZeroGPU. Every recorded scientific decision agreed; blanket
numerical identity did not. The control is neither a full 8,192-transition
reproduction nor a CPU-versus-GPU performance benchmark.

## License

Original Dataset material is licensed under the Creative Commons Attribution
4.0 International license. The complete terms are included in `LICENSE` and
are also available at https://creativecommons.org/licenses/by/4.0/.

Preferred attribution: “TiGa-RCE, independent researcher; founder and sole
operator of Brewster Jennings.” Third-party identifiers and factual references
are not relicensed by this Dataset.

## Integrity

`MANIFEST.sha256` binds every published Dataset file except the manifest
itself. The final HF Space must pin an exact Dataset commit rather than follow
the mutable `main` branch.

Status: public evidence Dataset. Publishing the Community Article remains a
separate final gate.