Extract section-typed full text for an arXiv paper (PAPI-06).

sections selects the projection:

• omitted or "all" (default): every canonical section — abstract, introduction, related_work, method, results, conclusion — at up to 3000 chars each, plus up to 5 captions.
• one section name: just that section.
• several names, comma-separated or as repeated params: just those, in canonical order. sections=method,results is roughly a third of the all payload, which matters when the caller is an LLM paying for every token of it.

Every response carries available_sections, so a caller always knows which sections this paper actually has. Requesting a section the paper does not contain returns 200 with requested_section_available: false, a note, and the sections that do exist — never an empty body.

section_provenance records how each section was typed, and low_confidence_sections names the ones that were inferred rather than read. Today that means only a method recovered by position: when no heading identified the method, the section between the introduction and the results is used, which a 50-paper hand audit put at ~50% accurate. Treat those as a hint and check the text before citing them.

Sources, tried in order: arXiv LaTeXML HTML (section-tagged, ~89% of the corpus), the LaTeX e-print tarball, then PDF text segmented by heading.

Rate limit: 10/minute per API key.