Research code for tensor-network similarity experiments.

uv.lock is committed; the install is reproducible against the exact versions used to produce every committed artifact under artifacts/.

Available <family> values: seed-convergence, curriculum-shift, language-similarity, subset-training.

The committed artifacts/cache/<family>/ is the canonical figure data (small .feather + .json files). uv run plot <family> reproduces the published figure from this cache directly. To regenerate the cache from scratch, delete it and re-run prepare.

• seed-convergence — cross-seed MNIST convergence (similarity + accuracy)

• curriculum-shift — 8-stage curriculum trajectory + pairwise heatmap

• language-similarity — pairwise functional similarity across pretrained language model checkpoints (pulls a log-spaced subsample of melephant/2l-bilinear-attn-normalised-v2 on first run; caches to _downloads/language-similarity/).

  Subsample size is N_STEPS=50 by default; tune via env. Each computed pair is appended to _progress.jsonl immediately, so an interrupted run resumes without recomputing. Rough budget on a single GPU after the first warm precompile: ~3.5 s per pair, so N×(N-1)/2 pairs at N=50→~30 min, N=75→~2.5 h, N=100→~5 h.

  N_STEPS=75 uv run prepare language-similarity # ~2.5 h uv run plot language-similarity # ~30 s once cache is built

• subset-training — Laurence-derived MNIST subset training across seeds

Keep durable figure code in src/figures/. EDA — anything that isn't producing data for the canonical figure — lives in workspaces/<user>/transient/.