TS Paper Latex

ts-paper-latex — assemble & compile the paper PDF (template-driven)

Turn the finished sections/<id>.tex + refs.bib + blueprint.json into a compiled main.pdf. This stage never authors or alters content — no renaming methods/datasets, no filling result numbers (proposal = blank cells stay blank). It is format + assembly + compile only.

Template-driven: assemble_paper.py reads template.json from the workdir and threads it through everything — section order + headings, table/figure caption position, keyword cap, whether to merge \cite (numeric templates only), and the preamble (the template's own main.tex.tmpl, filled by safe @@token@@ substitution) + which .sty/.cls + assets to copy. The default template is ts_iieta (two-column IIETA); neurips (single-column author-year) is also bundled and compiles through the same script — nothing here is hardcoded to TS. If no template.json is in the workdir it falls back to bundled ts_iieta (backward-compatible).

🔴 HARD RULE — NEVER fabricate a venue template. A template's style files (.sty/.cls) must come from exactly ONE of: (a) a user-provided template (the user drops a templates/<name>/ dir or points at the official files), or (b) the venue's OFFICIAL style files, fetched verbatim from the official source (the conference/journal style-file URL, e.g. NeurIPS media.neurips.cc/.../neurips_<year>.sty) and copied UNCHANGED. Do NOT hand-author, approximate, or "make it look like" a venue — a self-invented .sty gets margins/fonts/notice/line-numbers wrong and is unusable for submission. If a requested venue has neither a user-provided nor an obtainable official template, STOP and ask the user for the official files — never fabricate one. ⚠️ The currently-bundled ts_iieta and neurips styles are unofficial approximations ("official": false in their template.json); treat them as demo-only and replace with the venue's official .sty/.cls (or a user-provided template) before any real submission.

What's bundled (clean, copyright-safe assets you own)

• assets/ts_iieta.sty — our own two-column IIETA style (masthead with logo + blue band, 10pt Times, ALL-CAPS numbered sections, Figure N./Table N. captions, numeric [n] cites with sort&compress).
• assets/iieta_logo.png — the masthead logo (compiles fine if absent via \IfFileExists).
• scripts/assemble_paper.py — the deterministic backbone.

Run it

python scripts/assemble_paper.py <workdir>

The script: reads blueprint.json (title, keywords, abstract, section_order, optional authors/doi/dates/journal); for each sections/<id>.tex applies the post-processes and prepends the canonical ALL-CAPS heading; builds main.tex from the TS template; copies ts_iieta.sty + iieta_logo.png into the workdir; runs latexmk -pdf. It prints a JSON status {ok, compiled, exit, error_count, error_tail, sections, order, assets, template, workdir}.

Deterministic post-processes it applies (no LLM)

• Table captions ABOVE the tabular (IIETA convention) — both inline minipage and floating table/table*.
• Merge adjacent \cite runs into one (\cite{a} \cite{b} → \cite{a,b}) so they render as [1, 2] / [3-5] with the style's sort&compress.
• Canonical section headings — strips any heading the writer added and forces Introduction / Related Works / Methodology / Experimental Results / Discussion and Analysis / Conclusion, which the style renders as 1. INTRODUCTION …
• Strip heading numbers (\subsection{3.1 Foo} → Foo), table-width safety net (wraps a bare \begin{tabular} in \adjustbox), single-paragraph abstract merge.
• Keyword formatting — dedup, cap 6, comma-join into the \tsSetKeywords macro (empty → the line is omitted).
• The bundled ts_iieta.sty also loads inputenc[utf8] as a Unicode compile-net.

Compile + fix loop (quality gate — do not skip)

• The agent owns the ≤3-try fix loop; the script compiles once per call and only renders the verdict (no script-side loop). The script reports compiled (PDF exists AND error_count == 0 — it already trusts the log, not the exit code) and a structured error_count. Success also requires main.bbl resolved (no [?] citations). Treat error_count > 0 as a red gate that blocks done.
• Run with --backup on the FIRST attempt only so sections/ is snapshotted to sections.bak/ (rollback if a fix makes it worse) — do not re-pass --backup on later attempts or you overwrite the good baseline.
• If compiled:false or error_count > 0, read error_tail, then edit the offending sections/*.tex with minimal, syntax-only fixes (close unbalanced $/environments, escape stray & % # _ in text, fix mispaired \begin/\end, 1-10→1--10). Never change content, math semantics, citations, labels, or -- placeholders.
• Re-run. Bounded at ~3 attempts; if error_count goes UP, restore from sections.bak/ and report rather than thrash.
• One figure-specific error class: a File \figures/' not foundor a bad-bounding-box error is **not** a content bug — fix it by **re-running the figure stage** (regenerate the missing.pdf/.png`), never by editing section prose.
• Finally write logs/7_latex.io.md (INPUT: section files + refs.bib; DECISIONS: each compile error + the minimal fix; OUTPUT: main.pdf page count, error_count=0), and logs/index.md linking every stage's log.

Figures embed as vector PDF (no extra wiring)

Figures are referenced extension-less — \includegraphics{figures/<label>} — and the figure stage ships figures/<label>.pdf (editable vector) beside the kept figures/<label>.png. Both bundled .sty already \RequirePackage{graphicx}, so under pdflatex/latexmk the .pdf is embedded (its extension is preferred over .png) with no change to assemble_paper.py (it never parses \includegraphics or figure extensions). The vector is pre-rendered by cairosvg (in ts-figure-optimize's hybrid export), so there is no \includesvg/Inkscape/--shell-escape dependency. The run_gates.py all vector check confirms every figure has its .pdf sibling — a raster-only figure is a red gate.

Front matter / metadata

Authors/affiliation/email/DOI/dates default to clearly-marked placeholders ([AUTHORS TBD], 10.18280/ts.XXXXXX) for a proposal — that is correct; fill real values only if the user provides them in blueprint.json (authors, doi, dates, journal).

Done check

Report: page count, sections included, that captions are above tables, cites merged, keywords present, and that the citation linter (ts-paper-cite/scripts/citations_lint.py) passes with zero stubs/orphans.