table

Given a table number:

1. Find the producer. Check the project's own map first: an exhibit map (e.g. EXHIBIT_MAP.md), a generator script that stages and writes tables, or a Replication/ folder. If tables reach Overleaf through a generator, run the generator rather than the raw do-file, and respect its guards (e.g. an md5 check on the last-written file). Re-run any variant producers that share the table (suffix variants such as _imp).
2. Run it with the Stata console binary (run the stata-preflight checks first): on macOS that is the lowercase stata-se / stata-mp inside the app bundle, e.g. /Applications/Stata/StataSE.app/Contents/MacOS/stata-se -b do <file>. Check which edition is actually installed rather than assuming MP. Mac batch Stata chokes on spaces in the -b do path, so if any project directory has a space in its name, run it through a wrapper do-file in a space-free directory containing do "<quoted path>" (add a cd "<dir>" line first if the do-file uses relative paths).
3. Check the .log for r(###) errors and confirm its mtime is new; stop and report if not.
4. Clustering: keep what the do-file and the project's CLAUDE.md/memory specify. Project defaults override a global firm-level default (e.g. two-way firm + officer). Don't change it.
5. House style: read the project's CLAUDE.md/memory first. Common defaults:
   • Name the dependent variable in the table itself (a spanning "Dep. var.:" header, or a row when columns differ) — mlabels(, depvars) or a posthead line in esttab.
   • Notes describe construction only (data, sample, spec, FE, clustering, significance legend) and start "This table shows"; never describe or interpret results.
   • A column-number row; no N-firms rows unless asked.
   • Panels follow the paper's own layout. Some papers use one tabular per panel with \hline\hline and a centred italic title above each — that overrides the global booktabs rule for that paper.
   • esttab label gotcha: $F$ and backtick quotes vanish from labels()/postfoot() with no error. Grep the emitted .tex for the literal label text.
6. Recompile the document that \inputs the fragment. Don't assume a main.tex — a project may hold several standalone roots (the paper, a results appendix, a referee response), so compile the one that actually includes the fragment. Use tectonic (or latexmk if installed) with output in a temp dir so no PDFs land in the Overleaf-synced folder:

   tectonic -Z continue-on-errors --keep-logs -o "$TMP" <root>.tex

   Full drafts often carry LaTeX errors Overleaf ignores (e.g. float and floatrow both loaded, a subfigure clash), so plain tectonic halts on them. Compile a scratch copy of the Overleaf folder and compare the error count (grep -c '^!' "$TMP/<root>.log") with the draft's known baseline — only new errors matter. Never comment out a package to get a compile; it silently changes float and caption behaviour.
7. Check fit and placement for the affected page, and report actual measurements:
   • Overfull \hbox lines with their pt values.
   • Float too large warnings, and floats pushed to the end of the document.
   • Which page the table landed on: pdftotext -layout "$TMP/<root>.pdf" - and grep for the caption; pdfinfo for the page count. Then render that page and show the PNG, so the user can see it without opening the PDF:

   pdftoppm -f N -l N -r 110 -png "$TMP/<root>.pdf" "$TMP/table_p"

8. Only then commit, with a message naming the table, following the project's repo layout (some projects push only the code repo; logs and results stay local). Then sync to Overleaf (sync-overleaf).

Review loop ("move onto Table N")

When the user reviews exhibits one at a time: show the table as a PNG, list numbered proposed fixes, and wait for the user to pick ("do 1, 2 and 4"). Apply only those, re-render, and show it again. Don't carry a fix over to other tables unless asked.