publication-skills

name	Publication Skills
description	Research paper drafting, review, revision, and figure generation -- loaded when working on technical papers, white papers, or conference submissions
domain	pub
type	skill
user-invocable	false

Publications Domain

This project produces research papers in paper/. All publication work follows a structured draft-review-revise cycle with immutable version history and formal scoring.

State Machine

EMPTY --> DRAFTED --> REVIEWED --> REVISED --> REVIEWED --> ... --> READY

Commands map to state transitions:

Transition	Command	Input	Output
EMPTY --> DRAFTED	[[pub-draft]]	User interview + literature search	{thread}.1/
DRAFTED --> REVIEWED	[[pub-review]]	Version {N}/	{N}.review/review.md
REVIEWED --> REVISED	[[pub-revise]]	{N}/ + {N}.review/	{N+1}/
REVISED --> REVIEWED	[[pub-review]]	Version {N+1}/	{N+1}.review/review.md
READY --> AUDITED	[[pub-audit]]	Version {N}/	{N}.audit/audit.md
any state	[[pub-figures]]	Paper .tex	Missing figure scripts in figures/
portfolio	[[pub]]	Thread name or all	Assessment + parallel agent launch

Convergence criterion: review score >= 32/40 with 0 critical issues.

Naming Convention

{thread}.{N}

• thread: the paper topic name (e.g., group-moe)
• N: integer starting at 1, incremented by /pub-revise

Example: group-moe.3 is the 3rd revision of the Group-MoE paper.

Directory Layout

For this project, the thread name is group-moe. Paper versions live in paper/:

paper/
  outline.md                     # Working outline (paper_outline.md in docs/)
  group-moe.1/                   # Draft version 1
    paper.tex                    # LaTeX paper (standard article class)
    paper.pdf                    # Compiled PDF
    literature.md                # Literature review notes (internal)
    figures/                     # Python-generated figures (.py --> .pdf/.png)
    data/                        # Supporting scripts and results
  group-moe.1.review/            # Review (read-only sibling)
    review.md                    # Scored review report (markdown)
  group-moe.2/                   # Revision incorporating review
    ...                          # Same structure

Format Convention

• LaTeX (.tex --> .pdf): paper uses standard \documentclass{article} with academic packages (geometry, amsmath, booktabs, graphicx, hyperref). NOT sphere-patent.sty
• Markdown (.md): internal working documents -- literature.md, review.md
• Python figures (.py --> .pdf + .png): raw matplotlib for data plots
• Compile: pdflatex paper.tex (run twice for refs)

Key Principles

1. Literature review first. /pub-draft conducts a literature search before writing. Understand the landscape and position the contribution.
2. Immutable versions. Previous versions are never modified. The version history IS the revision trail.
3. Separation of concerns. Review is read-only (produces {N}.review/). Revision is separate (consumes {N}/ + {N}.review/, produces {N+1}/).
4. Converge, then submit. Cycle until review score >= 32/40 with 0 critical issues.

Commands

Command	Description
[[pub]]	Portfolio orchestrator: assess state of all papers, launch parallel agents
[[pub-draft]]	Interview + literature search + first-draft paper.tex (creates version 1)
[[pub-review]]	Read-only critic: 8-dimension scored review report
[[pub-revise]]	Consume draft + review, produce next version with all issues addressed
[[pub-audit]]	Fact-check: verify citations, numbers, equations, reproducibility claims
[[pub-figures]]	Batch-generate missing figures for a paper

Checkpointing Protocol

Long-running skills (/pub-draft, /pub-revise, /pub-figures) write _progress.json in the output directory. On retry, completed phases are skipped.

Skill	Checkpointed Phases
/pub-draft	interview, literature_search, paper_tex, compile_pdfs
/pub-revise	read_inputs, paper_tex, literature_update, compile_pdfs, self_check
/pub-figures	analysis, then per-figure: fig_N_script, fig_N_run, fig_N_verify
/pub-review	Not checkpointed (single output file)
/pub-audit	Not checkpointed (single output file)

Key References

• docs/paper_outline.md -- comprehensive paper outline with all experimental results
• docs/arithmetic_experiment.md -- S_2 complement transfer results
• docs/ternary_experiment.md -- S_3 complement transfer results
• data/ -- JSON files with all experimental outputs (gitignored)
• scripts/ -- experiment scripts (source of all reported results)