[ParsaVictor]

Turns docs/research/paper-draft.md into an arXiv-ready LaTeX paper in paper/. No benchmarks or builds were run. paper-draft.md and contributions-log.md are unchanged.

Files

• paper/main.tex: article class. One column by default; make twocol builds the two-column version from the same source.
• paper/tables/*.tex: seven tables. They cover the Lite-274 head-to-head (same subset and metric only), all holdouts (BM25 vs engine, file and function level), the dense short-list vote, the ablation, the plain-language holdouts, cost, and negative results. Each table's header comment names the log section its numbers come from.
• paper/refs.bib: 39 entries, all of them cited. The author lists and titles of the 28 arXiv entries come from their arXiv abstract pages, and each entry carries its eprint id. Venues follow the related-work file (docs: related work and competitor table (session 19) #146).
• paper/Makefile: targets all, twocol, tectonic, check, todo (fails while any \todo is left), arxiv (packs the .bbl with the sources) and clean.
• paper/check.py: static checks for a machine with no TeX installed. It checks brace and environment balance, \input files, cite keys, labels and ASCII-only source, and it lists TODOs.
• paper/REVIEWER-RISKS.md: the 10 hardest reviewer objections, with our current answer to each and the experiment that is still missing.

What the paper adds beyond the draft:

• The session-19 results:
  • SWE-Gym as a third holdout;
  • the learned reranker, rejected;
  • the dense short-list vote: Lite-274 function Acc@1/5/10 goes from 0.223/0.460/0.540 to 0.252/0.485/0.555 (paired change at @5 +0.026 [0.004, 0.051], 9 gained / 2 lost, p = 0.065, reported as such).
• The full competitor table from docs: related work and competitor table (session 19) #146.

Positioning: the paper states up front that the engine is below CodeRankEmbed and SweRankEmbed-Small (0.664 / 0.631) and below the LLM agents. The claim is the no-model, CPU-only, zero-token, holdout corner of the trade-off.

Still open (make check lists them):

• The Verified number for the dense vote (it is being measured).
• Affiliation and co-authors.
• A Zenodo DOI.
• Funding.
• The wording of the LLM-assistance disclosure.

Not verified: the source has not been compiled, because there is no pdflatex or tectonic on this machine. python paper/check.py passes. Run make once wherever TeX is available.

🤖 Generated with Claude Code