repro-ingest

name	repro-ingest
description	Read an academic paper and extract its method, datasets, hyperparameters, compute needs, and every reproducible quantitative result into structured claims. Use as the first step of reproducing a paper, after init_reproduction.

Ingest a paper into claims

Goal: turn the paper into (a) a self-contained .repro/paper.md and (b) a registered claim for every quantitative result you intend to reproduce.

Get the text

• arXiv id or arXiv URL → prefer the HTML version (https://arxiv.org/abs/<id> and the ar5iv/HTML full text) over the PDF; it is far easier to parse tables. Fall back to the PDF if HTML is unavailable.
• DOI / publisher URL → fetch the landing page; find an open-access PDF/HTML.
• Local PDF → read it directly.
• Save anything you download under .repro/sources/.

Fill in .repro/paper.md

Capture, concisely but completely:

• Summary — what the paper does and its main claims.
• Method — the algorithm/model/architecture, enough to implement it.
• Datasets — names, sizes, splits, preprocessing, and where to obtain them.
• Hyperparameters & training setup — optimizer, lr schedule, batch size, epochs/steps, seeds, regularization, anything needed to reproduce.
• Compute requirements — hardware, wall-clock, and rough cost. Flag early if reproduction needs resources the user may not have.

Identify claims

Go through the paper's tables, key figures, and ablations. For each quantitative result you plan to reproduce, call register_claim with:

• a stable id (e.g. table2-cifar10-acc),
• metric_name, reported_value, unit,
• reported_std when the paper gives the result as mean ± σ over seeds/folds — this lets agreement be judged by statistical consistency, not a fixed tolerance,
• direction (higher/lower = better) so beating the paper still counts,
• source_ref (e.g. "Table 2, row CIFAR-10"), dataset, and any notes.

Guidance:

• Prioritize headline results — the main table and the central figure — over exhaustive ablation grids. Register the most load-bearing claims first; you can add more later.
• Separate reproducible from qualitative. Numbers in tables are claims. Qualitative statements ("our method is more robust") are not — note them in paper.md but don't register them.
• One number per claim. A table row with 4 columns is 4 claims if you intend to reproduce all 4.
• Note the provenance when a reported number depends on pretrained weights, external data, or a specific seed — it affects how you reproduce it.

When done, hand back to repro-create (next: gap-analysis if a repo exists, otherwise planning).