Publication Cycle Skill

Use $PUB_DIR/backflow/backflow_inventory.json and each paper's PIPELINE.md to determine coverage.

1b. Classify existing papers

For each paper in $PUB_DIR/2026_*/:

1. Read P4_EDITORIAL_REVIEW*.md files to get the latest verdict
2. Classify into tiers:

1c. Output: Action Plan

Produce a prioritized action plan:

## New Extractions
- [section] → [proposed paper slug] → [target journal]

## Paper Queue (priority order)
1. [paper] — Tier [X] — [next action]
2. ...

## Backflow Queue
- [ACCEPT paper] → [core section] — [status]

Phase 2: EXECUTE — Review+Fix Pipeline

For each paper, execute in order:

Step 2.1: Codex General Review + Self-Fix

python $PUB_DIR/codex_fix.py --paper $PAPER_DIR --review-text "
Perform a general editorial review of this paper. Check:
1. Mathematical correctness of all theorems and proofs
2. Bibliography completeness (all citations resolve)
3. Author field populated
4. No orphaned or unreferenced files
5. Cross-references all resolve
Fix every issue you find."

Step 2.2: Codex Targeted Review + Fix

python $PUB_DIR/codex_fix.py --paper $PAPER_DIR --review-text "
Perform a targeted review for [TARGET_JOURNAL]. Check:
1. Writing style matches journal conventions
2. Novelty is clearly stated and not oversold
3. Related work section covers the relevant literature
4. Proof completeness — every statement has a proof
5. Abstract and introduction are publication-ready
Fix every issue you find."

Step 2.3: ChatGPT Review (via Oracle)

python $PUB_DIR/oracle_dispatch.py --paper $PAPER_DIR --prompt editorial_review

Wait for oracle response. Parse review from oracle/done/.

Step 2.4: Codex Fix from ChatGPT Feedback

python $PUB_DIR/codex_fix.py --paper $PAPER_DIR --review [chatgpt_review_file]

Step 2.5: Claude Deep Review

Launch pub-editorial agent for deep mathematical verification.

Step 2.6: Codex Fix from Claude Feedback

python $PUB_DIR/codex_fix.py --paper $PAPER_DIR --review [claude_review_file]

Step 2.7: ChatGPT Journal-Targeted Acceptance Gate

python $PUB_DIR/oracle_dispatch.py --paper $PAPER_DIR --prompt acceptance_gate

If ACCEPT → move to Phase 3 (backflow + P7). If not ACCEPT → return to Step 2.4 with new feedback.

Tier shortcuts:

• Tier A: Skip all steps, go to Phase 3
• Tier B: Skip 2.1-2.2, start at 2.3
• Tier C: Skip 2.1, start at 2.2
• Tier D/E/F: Full pipeline from 2.1

Phase 3: BACKFLOW + SUBMISSION

For ACCEPT papers:

1. Run python $PUB_DIR/backflow.py inject --execute
2. Update PIPELINE.md backflow table
3. Prepare P7 submission pack if all gates passed

Phase 4: SPLIT CHECK

After each review round, check if the reviewer identified:

• Sections that should be separate papers
• Results that don't fit the target journal's scope
• Content that would strengthen a different paper

If splitting is recommended:

1. Create new paper directory
2. Move relevant .tex content
3. Add to paper queue at appropriate tier

Automation Tools Reference

Token Budget Strategy

1. Codex (cheapest per fix): Do ALL editing. 2-3 rounds per paper.
2. ChatGPT (free): Independent validation + acceptance gate.
3. Claude (most expensive): Deep math verification only. 1 round per paper.

Parallel Execution

While waiting for ChatGPT oracle (slow, browser-based):

• Run Codex on other papers
• Run backflow for ACCEPT papers
• Run Lean sync checks

While waiting for Codex (moderate):

• Prepare ChatGPT prompts for next papers
• Update PIPELINE.md files