Papers¶

papers/registry.toml is the source of truth for papers that directly guide current research or implementation. Each entry owns one publication citation, one exact arXiv revision, and the source license shown by arXiv.

tracked registry
      |
      +--> publication citation
      +--> arXiv /abs/<exact revision>
      +--> arXiv /pdf/<exact revision>
      +--> arXiv /src/<exact revision>
                         |
                         v
              papers/_cache/  (ignored)

List the registry, print BibTeX, or fetch the rendered paper and TeX archive:

uv run --locked python -m papers.run --list
uv run --locked python -m papers.run --cite i-jepa graph-jepa
uv run --locked python -m papers.run i-jepa graph-jepa

With no paper names, each command selects the whole registry. Fetches are bounded, atomic, and stored under the paper key and exact arXiv revision. The command prints a SHA-256 digest for every local file.

The current sources are I-JEPA v3, published at CVPR 2023, and Graph-JEPA v3, published in TMLR, plus GINE v3, published at ICLR 2020, and V-JEPA v1, and TS-JEPA v1, presented at the NeurIPS 2024 time-series workshop.

Source rule¶

Add a paper only when a live research question or implementation uses it. Pin an explicit vN, copy citation metadata from the publication venue, and record the license linked by that arXiv revision. Cite the publication in prose and name the exact arXiv revision when an equation, algorithm, loss, or evaluation protocol affects the work.

Read the PDF or TeX before making an implementation claim. Do not commit paper copies or extracted source; the cache is reproducible and replaceable. A source license may permit less than redistribution, which is another reason the tracked boundary contains metadata rather than paper bodies.