| name | publish |
| description | Use when a paper draft exists and needs to be prepared for venue submission or arxiv preprint |
| complexity | high |
| model-minimum | gpt-5 |
| disable-model-invocation | false |
| allowed-tools | ["Read","Grep","Glob","WebSearch","WebFetch","Bash","Write","Edit"] |
| argument-hint | [project name or paper path] |
/publish
You are preparing a research paper for submission. Your job is to take an existing draft through a structured publication pipeline: verify completeness, check citations, format for the target venue, create figures, anonymize, and produce a submission-ready artifact.
The argument is a project name (reads publications/ directory) or a direct path to a paper file.
Step 0: Locate and assess the paper
- Find the paper draft. If given a project name, look in
projects/<name>/publications/. If given a path, read it directly.
- Read the paper and identify:
- Current state (draft, formatted, submission-ready)
- Target venue (check for venue selection experiment or task notes)
- Deadline (check TASKS.md for deadline notes)
- Content completeness (are all sections filled in?)
- Read the project README and TASKS.md for context on what the paper covers.
Step 0.5: Synthesis pre-flight audit (mandatory)
Before assessing content, verify that the upstream data the paper builds on is actually correct. Follow the procedure at docs/sops/synthesis-preflight-audit.md:
- Enumerate upstream sources — list every experiment, analysis, and literature note the paper cites
- Flag provisional data — check
status: and data_quality: fields in EXPERIMENT.md frontmatter
- Verify key numerical claims — for numbers that appear in the abstract, tables, or drive conclusions, trace to the computational source (script + data file) and re-verify. Do NOT accept numbers copied from text.
- Verify literature citations — check
Verified: field in literature notes; fetch URLs for any unverified citations
- Check cross-experiment comparisons — verify denominators match and are explicitly stated
- Gate decision — if any key claim is unverifiable or mismatched, fix before proceeding to Step 1
This step catches the most common synthesis failure: treating prior session outputs as verified ground truth when they may contain contaminated data, stale numbers, or fabricated citations (7 documented incidents — see projects/akari/feedback/feedback-incorrect-experiment-contamination-analysis-2026-03-01.md).
Step 1: Content audit
Check the paper against these requirements:
Structure:
- Abstract (concise, states the problem, approach, key results, and conclusion)
- Introduction (motivation, contributions, paper organization)
- Related work (positions against prior art, cites relevant work)
- Method/approach (reproducible description)
- Results (quantitative with proper metrics)
- Discussion (limitations, implications, future work)
- Conclusion
Provenance:
- Every numerical claim has a source (script + data file, or inline arithmetic)
- Every citation has a verified URL/DOI (check
verified field in literature notes)
- No claims sourced from parametric memory alone
Missing content flags:
- Sections with placeholder text or TODOs
- Results referenced but not presented
- Figures referenced but not created
- Related work entries without verified citations
Report gaps as a checklist. If critical gaps exist, stop and list them before proceeding.
Step 2: Citation verification
Run /audit-references on the paper or manually verify:
- Every in-text citation maps to a reference entry
- Every reference entry has a verified URL/DOI
- No orphan references (listed but never cited)
- No orphan citations (cited but not in reference list)
If unverified citations exist, flag them. For submissions, all citations must be verified.
Step 3: Venue formatting
Based on the target venue:
-
Identify format requirements:
- Page limit (content pages vs total including refs)
- Template/style file (LaTeX class, author kit)
- Anonymization requirements (double-blind, single-blind, open)
- Supplemental materials policy
- Figure/table formatting guidelines
-
Check tool availability:
- LaTeX compiler available? (
which pdflatex xelatex lualatex tectonic 2>/dev/null)
- Plotting tools? (
python -c "import matplotlib" 2>/dev/null)
- If tools are missing, note this as a blocker and file a tool-access request to
APPROVAL_QUEUE.md if not already filed.
-
If tools are available, format the paper:
- Convert markdown to LaTeX using the venue template
- Create figures from data (tables → charts where appropriate)
- Verify page count fits venue limits
- Generate PDF and check formatting
-
If tools are NOT available:
- Prepare the markdown content to be as close to submission-ready as possible
- Create figure specifications (what each figure should show, data source, chart type)
- Document the exact formatting steps needed when tools become available
- Ensure all content fits estimated page count
Step 4: Anonymization (if required)
For double-blind venues:
- Remove author names and affiliations
- Replace self-references with anonymous placeholders ("We previously showed [Anonymous, 2026]")
- Remove repository URLs that identify authors
- Check acknowledgments section
- Remove or anonymize any identifying information in supplemental materials
- Search for project name, author names, and repo URLs throughout the document
For open review venues: skip this step.
Step 5: Self-review checklist
Before declaring submission-ready, verify:
Step 6: Produce submission artifacts
Create the submission directory:
projects/<project>/publications/<paper-name>/<venue>/
paper.tex (or paper.md if no LaTeX)
figures/
supplemental/
submission-checklist.md
The submission-checklist.md records:
- Venue name and deadline
- Submission format (PDF, LaTeX source, etc.)
- Anonymization status
- Tool access status (what's available, what's missing)
- Outstanding items before submission
- Venue-specific requirements checklist
Output
After completing all steps, report:
- Paper state (ready, near-ready with gaps listed, blocked)
- Outstanding items (as a numbered list with effort estimates)
- Any approval items filed
Common Rationalizations
| Excuse | Reality |
|---|
| "The citation is probably correct, I'll skip verification" | Hallucinated citations are the #1 quality risk. Verify every citation mechanically. |
| "We can fix formatting later" | Formatting reveals content gaps (figures that don't exist, results that don't fit). Format early. |
| "Anonymization is trivial" | One missed self-reference or repo URL can reveal authorship. Use systematic search. |
| "The page limit is a soft guideline" | Venues desk-reject papers that exceed page limits. Treat limits as hard constraints. |
Commit
Follow docs/sops/commit-workflow.md. Commit message: publish: <paper-name> — <venue> submission preparation