Workflow 3: Paper Writing Pipeline

Orchestrate a complete paper writing workflow for: $ARGUMENTS

Overview

This skill chains five sub-skills into a single automated pipeline:

/paper-plan → /paper-figure → /paper-write → /paper-compile → /research-refine
  (outline)     (plots)        (LaTeX)        (build PDF)       (review & polish ×2)

Each phase builds on the previous one's output. The final deliverable is a polished, reviewed paper/ directory with LaTeX source and compiled PDF.

In this hybrid pack, the pipeline itself is unchanged, but paper-plan and paper-write use Orchestra-adapted shared references for stronger story framing and prose guidance.

Constants

• VENUE = ICLR — Target venue. Options: ICLR, NeurIPS, ICML, CVPR, ACL, AAAI, ACM, IEEE_JOURNAL (IEEE Transactions / Letters), IEEE_CONF (IEEE conferences), or any discipline-specific venue name (e.g., 法学研究, AER, NEJM, American Sociological Review). When a discipline-specific venue is used, the pipeline reads PRISM_PROFILE.md for formatting rules. Affects style file, page limit, citation format.
• MAX_IMPROVEMENT_ROUNDS = 2 — Number of review→fix→recompile rounds in the improvement loop.
• REVIEWER_MODEL = gpt-5.4 — Model used via Codex MCP for plan review, figure review, writing review, and improvement loop.
• AUTO_PROCEED = true — Auto-continue between phases. Set false to pause and wait for user approval after each phase.
• HUMAN_CHECKPOINT = false — When true, the improvement loop (Phase 5) pauses after each round's review to let you see the score and provide custom modification instructions. When false (default), the loop runs fully autonomously. Passed through to /research-refine.
• ILLUSTRATION = figurespec — Architecture/illustration generator for Phase 2b: figurespec (default, deterministic JSON→SVG via /figure-spec, best for architecture/workflow/topology), gemini (AI-generated via /paper-illustration with its default Gemini backend, best for qualitative method illustrations; needs GEMINI_API_KEY), codex-image2 (AI-generated via /paper-illustration — backend: codex through the local Codex native image bridge — no external API key, uses your ChatGPT Plus/Pro quota; experimental), mermaid (Mermaid syntax via /mermaid-diagram, free, best for flowcharts), or false (skip Phase 2b, manual only).

Override inline: /paper-writing "NARRATIVE_REPORT.md" — venue: NeurIPS, illustration: gemini, human checkpoint: true IEEE example: /paper-writing "NARRATIVE_REPORT.md" — venue: IEEE_JOURNAL

Inputs

This pipeline accepts one of:

1. NARRATIVE_REPORT.md (best) — structured research narrative with claims, experiments, results, figures
2. Research direction + experiment results — the skill will help draft the narrative first
3. Existing PAPER_PLAN.md — skip Phase 1, start from Phase 2

The more detailed the input (especially figure descriptions and quantitative results), the better the output.

Pipeline

Phase 0: Assurance Setup

Resolve the active assurance level and persist it so Phase 6's external verifier reads the same value. Run once at pipeline start, before Phase 1.

Resolution order (first match wins):

1. Explicit — assurance: draft | submission in $ARGUMENTS
2. Derived from — effort:
   • lite / balanced → draft (default, zero change from current behavior)
   • max / beast → submission
3. Default: draft

Action:

mkdir -p paper/.prism
echo "<resolved-level>" > paper/.prism/assurance.txt   # draft or submission

What each level does downstream:

• draft — Existing behavior. Audits run only when their content detector matches (Phase 4.5 / 4.7 / 5.5 / 5.8). Missing artifacts are non-blocking. Silent-skip allowed.
• submission — The three mandatory audits (proof-checker, paper-claim-audit, citation-audit) are treated as load-bearing gates. Each sub-audit must emit its JSON artifact (PASS / WARN / FAIL / NOT_APPLICABLE / BLOCKED / ERROR) — never silent-skip. Phase 6 runs tools/verify_paper_audits.sh; a non-zero exit blocks the Final Report.

Escape hatch: a user wanting the old "beast = depth-only, no audit gate" can pass — effort: beast, assurance: draft explicitly. Legal but discouraged for actual submissions. See shared-references/assurance-contract.md for the full contract.

Announce the resolved level in-line before Phase 1:

📋 Assurance: <level> (derived from effort: <effort>)
   <either "current behavior, no audit gate" OR "mandatory audits gated by tools/verify_paper_audits.sh">

Phase 1: Paper Plan

Invoke /paper-plan to create the structural outline:

/paper-plan "$ARGUMENTS"

What this does:

• Parse NARRATIVE_REPORT.md for claims, evidence, and figure descriptions
• Build a Claims-Evidence Matrix — every claim maps to evidence, every experiment supports a claim
• Design section structure (5-8 sections depending on paper type)
• Plan figure/table placement with data sources
• Scaffold citation structure
• GPT-5.4 reviews the plan for completeness

Output: PAPER_PLAN.md with section plan, figure plan, citation scaffolding.

Checkpoint: Present the plan summary to the user.

📐 Paper plan complete:
- Title: [proposed title]
- Sections: [N] ([list])
- Figures: [N] auto-generated + [M] manual
- Target: [VENUE], [PAGE_LIMIT] pages

Shall I proceed with figure generation?

• User approves (or AUTO_PROCEED=true) → proceed to Phase 2.
• User requests changes → adjust plan and re-present.

Phase 2: Figure Generation

Invoke /paper-figure to generate data-driven plots and tables:

/paper-figure "PAPER_PLAN.md"

What this does:

• Read figure plan from PAPER_PLAN.md
• Generate matplotlib/seaborn plots from JSON/CSV data
• Generate LaTeX comparison tables
• Create figures/latex_includes.tex for easy insertion
• GPT-5.4 reviews figure quality and captions

Output: figures/ directory with PDFs, generation scripts, and LaTeX snippets.

Scope: paper-figure covers data plots and comparison tables. Architecture diagrams, pipeline figures, and method illustrations are handled in Phase 2b below.

Phase 2b: Architecture & Illustration Generation

Skip this step entirely if illustration: false.

If the paper plan includes architecture diagrams, pipeline figures, audit cascades, or method illustrations, invoke the appropriate generator based on the illustration parameter:

When illustration: figurespec (default) — invoke /figure-spec:

/figure-spec "[architecture/workflow description from PAPER_PLAN.md]"

• Deterministic JSON → SVG vector rendering (editable, reproducible)
• Best for: system architecture, workflow pipelines, audit cascades, layered topology
• Output: figures/*.svg + figures/*.pdf (via rsvg-convert) + figures/specs/*.json
• No external API, runs fully local

When illustration: gemini — invoke /paper-illustration:

/paper-illustration "[method description from PAPER_PLAN.md or NARRATIVE_REPORT.md]"

• Claude plans → Gemini optimizes → Nano Banana Pro renders → Claude reviews (score ≥ 9)
• Best for: qualitative method illustrations, natural-style diagrams, result grids
• Output: figures/ai_generated/*.png
• Requires GEMINI_API_KEY environment variable

When illustration: mermaid — invoke /mermaid-diagram:

/mermaid-diagram "[method description from PAPER_PLAN.md]"

• Generates Mermaid syntax diagrams (flowchart, sequence, class, state, etc.)
• Best for: lightweight flowcharts, state machines, simple sequence diagrams
• Output: figures/*.mmd + figures/*.png
• Free, no API key needed

When illustration: codex-image2 — invoke /paper-illustration with the codex backend:

/paper-illustration "[method description from PAPER_PLAN.md or NARRATIVE_REPORT.md]" — backend: codex

• Claude plans → Codex native image generation renders → Claude reviews (same multi-stage workflow as gemini, different renderer)
• Best for: users who want a GPT-image-style renderer without needing GEMINI_API_KEY; uses your existing Codex / ChatGPT Plus/Pro quota
• Output: figures/ai_generated/figure_final.png + latex_include.tex + review_log.json (emitted via tools/paper_illustration_image2.py finalize)
• Prerequisites (beyond Prism's standard Claude Code + Codex coexistence): the local Codex app-server must be signed in (codex debug app-server send-message-v2 "ping" succeeds), and the dedicated MCP bridge must be registered — see mcp-servers/codex-image2/README.md for the one-time claude mcp add command. Run python3 tools/paper_illustration_image2.py preflight --workspace . to confirm before relying on this path.
• Experimental: this renderer shells through the Codex debug app-server, which Codex documents as an unstable surface. Prefer figurespec or gemini for production submission flows until codex-image2 stabilizes.

When illustration: false — skip entirely. All non-data figures must be created manually (draw.io, Figma, TikZ) and placed in figures/ before Phase 3.

Choosing the right mode:

• Formal architecture / workflow / topology figures → figurespec (default)
• Method concept illustrations with natural style, have GEMINI_API_KEY → gemini
• Method concept illustrations, prefer ChatGPT Plus/Pro quota over Gemini key → codex-image2
• Quick flowchart / state machine → mermaid
• Full manual control → false

These are complementary, not mutually exclusive: you can run multiple generators for different figures in the same paper by re-invoking with different illustration overrides.

Checkpoint: List generated vs manual figures.

📊 Figures complete:
- Data plots (auto, Phase 2): [list]
- Architecture/illustrations (auto, Phase 2b, mode=<illustration>): [list]
- Manual (need your input): [list]
- LaTeX snippets: figures/latex_includes.tex

[If manual figures needed]: Please add them to figures/ before I proceed.
[If all auto]: Shall I proceed with LaTeX writing?

Phase 3: LaTeX Writing

Invoke /paper-write to generate section-by-section LaTeX:

/paper-write "PAPER_PLAN.md"

What this does:

• Write each section following the plan, with proper LaTeX formatting
• Insert figure/table references from figures/latex_includes.tex
• Build references.bib from citation scaffolding
• Clean stale files from previous section structures
• Automated bib cleaning (remove uncited entries)
• De-AI polish (remove "delve", "pivotal", "landscape"...)
• GPT-5.4 reviews each section for quality

Output: paper/ directory with main.tex, sections/*.tex, references.bib, math_commands.tex.

Checkpoint: Report section completion.

✍️ LaTeX writing complete:
- Sections: [N] written ([list])
- Citations: [N] unique keys in references.bib
- Stale files cleaned: [list, if any]

Shall I proceed with compilation?

Phase 4: Compilation

Invoke /paper-compile to build the PDF:

/paper-compile "paper/"

What this does:

• latexmk -pdf with automatic multi-pass compilation
• Auto-fix common errors (missing packages, undefined refs, BibTeX syntax)
• Up to 3 compilation attempts
• Post-compilation checks: undefined refs, page count, font embedding
• Precise page verification via pdftotext
• Stale file detection

Output: paper/main.pdf

Checkpoint: Report compilation results.

🔨 Compilation complete:
- Status: SUCCESS
- Pages: [X] (main body) + [Y] (references) + [Z] (appendix)
- Within page limit: YES/NO
- Undefined references: 0
- Undefined citations: 0

Shall I proceed with the improvement loop?

Phase 4.5: Proof Verification (theory papers only)

Skip this phase if the paper contains no theorems, lemmas, or proofs.

if paper contains \begin{theorem} or \begin{lemma} or \begin{proof}:
    Run /proof-checker "paper/"
    This invokes GPT-5.4 xhigh to:
    - Verify all proof steps (hypothesis discharge, interchange justification, etc.)
    - Check for logic gaps, quantifier errors, missing domination conditions
    - Attempt counterexamples on key lemmas
    - Generate PROOF_AUDIT.md with issue list + severity

    If FATAL or CRITICAL issues found:
        Fix before proceeding to improvement loop
    If only MAJOR/MINOR:
        Proceed, improvement loop may address remaining issues
else:
    skip — no proofs, no action

Phase 4.7: Paper Claim Audit

Skip if no result files exist (e.g., survey/position papers with no experiments).

if results/*.json or results/*.csv or outputs/*.json exist:
    Run /paper-claim-audit "paper/"
    Fresh zero-context reviewer compares every number in the paper
    against raw result files. Catches rounding inflation, best-seed
    cherry-pick, config mismatch, delta errors.

    If FAIL:
        Fix mismatched numbers before improvement loop
    If WARN:
        Proceed, but flag for manual verification
else:
    skip — no experimental results to verify

Phase 4.8: Prose Lint (deterministic, model-free)

Before the model-driven improvement loop, run the deterministic text-review lint over the prose so the loop spends its review budget on substance, not hygiene:

python <PRISM_REPO>/tools/text_review.py paper/sections/*.md --severity warning

It catches the recurring humanities/social-science defects regex can decide: vague or future-dated citations (R1-*), over-assertion (R3-01), terminology inconsistency (T-01), GB/T 7714 / footnote / punctuation slips (F-*), register and self-reference drift (S-*), missing structural elements (ST-*), and the argument-logic flags (L-*: assert-without-argue, list-without-advance, quotation-without-analysis, unsupported intensifiers). Fix error-level issues before Phase 5; feed surviving L-* flags to /research-refine as targeted revision notes. Advisory unless assurance: submission, where unresolved error-level issues block the Final Report alongside the other gates.

Phase 5: Auto Improvement Loop

Invoke /research-refine to polish the paper:

/research-refine "paper/"

What this does (2 rounds):

Round 1: GPT-5.4 xhigh reviews the full paper → identifies CRITICAL/MAJOR/MINOR issues → Claude Code implements fixes → recompile → save main_round1.pdf

Round 2: GPT-5.4 xhigh re-reviews with conversation context → identifies remaining issues → Claude Code implements fixes → recompile → save main_round2.pdf

Typical improvements:

• Fix assumption-model mismatches
• Soften overclaims to match evidence
• Add missing interpretations and notation
• Strengthen limitations section
• Add theory-aligned experiments if needed

Output: Three PDFs for comparison + PAPER_IMPROVEMENT_LOG.md.

Format check (included in improvement loop Step 8): After final recompilation, auto-detect and fix overfull hboxes (content exceeding margins), verify page count vs venue limit, and ensure compact formatting. Location-aware thresholds: any main-body overfull blocks completion regardless of size; appendix overfulls block only if >10pt; bibliography overfulls block only if >20pt.

Phase 5.5: Final Paper Claim Audit (MANDATORY submission gate)

After /research-refine finishes, rerun /paper-claim-audit before the final report whenever the paper contains numeric claims and machine-readable raw result files exist.

Use the same detectors as Phase 4.7:

• numeric-claim regex over paper/main.tex and paper/sections/*.tex
• raw-evidence file search in results/, outputs/, experiments/, and figures/ for .json, .jsonl, .csv, .tsv, .yaml, or .yml

This phase is mandatory if both detectors are positive. It blocks the final report. If numeric claims exist but no raw result files are found, stop and warn the user before declaring the paper complete. If no numeric claims exist, skip.

NUMERIC_CLAIMS=$(rg -n -e '[0-9]+(\.[0-9]+)?\s*(%|\\%|±|\\pm|x|×)' \
  -e '(accuracy|BLEU|F1|AUC|mAP|top-1|top-5|error|loss|perplexity|speedup|improvement)' \
  paper/main.tex paper/sections 2>/dev/null || true)

RAW_RESULT_FILES=$(find results outputs experiments figures -type f \
  \( -name '*.json' -o -name '*.jsonl' -o -name '*.csv' -o -name '*.tsv' -o -name '*.yaml' -o -name '*.yml' \) 2>/dev/null | head -200)

if [ -n "$NUMERIC_CLAIMS" ] && [ -n "$RAW_RESULT_FILES" ]; then
    Run /paper-claim-audit "paper/"
    If FAIL:
        Fix mismatched numbers before the final report
elif [ -n "$NUMERIC_CLAIMS" ]; then
    Stop and warn: the paper contains numeric claims but no raw evidence files were found
fi

Empirical motivation: in our April 2026 NeurIPS run, the final paper claimed w ∈ {0,1,2,3} for the width-tradeoff experiment but the raw JSON had w ∈ {0,1,2,3,4,5}. The crossing-point tolerance was claimed as 0.05% but the actual relative error was 0.0577%. Both were caught only after manual paper-claim-audit invocation in the final round; the improvement loop did not detect them.

Phase 5.8: Citation Audit (submission gate)

After the final paper-claim-audit passes, run /citation-audit to verify every \cite{...} along three axes: existence, metadata correctness, and context appropriateness. This is the fourth and final layer of the evidence-and-claim assurance stack (paper-claim-audit → argument-stress-test → paper-claim-audit → citation-audit).

if paper/references.bib (or paper.bib) exists and contains entries cited from sec/*.tex:
    Run /citation-audit "paper/"
    Fresh cross-family reviewer (gpt-5.4 via Codex MCP) with web/DBLP/arXiv lookup
    verifies each entry:
      (i)   EXISTENCE — paper resolves at claimed arXiv ID / DOI / venue
      (ii)  METADATA — author names, year, venue, title match canonical sources
      (iii) CONTEXT — cited paper actually establishes the claim it supports

    Output:
      - CITATION_AUDIT.md (human-readable per-entry verdict report)
      - CITATION_AUDIT.json (machine-readable verdict ledger)
      - Per-entry verdicts: KEEP / FIX / REPLACE / REMOVE

    If any REPLACE or REMOVE verdicts:
        Surface to user for human approval — never auto-modify content claims
    If only FIX verdicts (metadata corrections):
        Apply with user confirmation, then recompile
    If all KEEP:
        Pass — bibliography clean for submission
else:
    skip — no bib file or no citations

Why this is the most diagnostic of the four audit layers: wildly fake citations are easy to spot. The dangerous failure mode is a real paper used to support a claim it does not actually establish (wrong-context citations) — these slip past metadata-only checks and damage submission credibility. Run cost is wall-clock heavy (web lookup per entry); run once per submission, not per save.

Empirical motivation: in our April 2026 Prism technical-report run, two real papers (madaan2023selfrefine, liu2023reviewergpt) were cited in contexts they did not actually support, and one entry (hidden2025aiscientistpitfalls) had author = "Anonymous" because the metadata had not been resolved. None were caught by the improvement loop or numeric claim audit; only fresh web-lookup review surfaced them.

Phase 6: Final Report

Phase 6.0 — Submission Gate

Before writing the Final Report, resolve the active assurance level. This uses the same derivation rule as Phase 0 so a run where Phase 0 was skipped or its write failed cannot silently downgrade a beast / max / — assurance: submission invocation back to draft.

Resolution at the gate (re-derive; do not trust .prism/assurance.txt alone):

1. Parse $ARGUMENTS for an explicit — assurance: draft | submission or an — effort: lite | balanced | max | beast directive.
2. Derive the expected level:
   • explicit assurance: wins
   • else lite / balanced → draft, max / beast → submission
   • else draft
3. Read paper/.prism/assurance.txt. If the file is missing, write it now with the derived level.
4. If the file's value disagrees with the derived level (e.g. file says draft but $ARGUMENTS says beast), overwrite the file with the derived level and surface a one-line warning in-chat: ⚠️ .prism/assurance.txt was draft but $ARGUMENTS says submission; overriding.
5. Use the re-derived level as authoritative for the rest of Phase 6.

# Final authoritative value, written and read from the same source
ASSURANCE=<derived-from-$ARGUMENTS>        # draft | submission
mkdir -p paper/.prism
echo "$ASSURANCE" > paper/.prism/assurance.txt

If ASSURANCE=draft, skip directly to the Final Report template below — current behavior, no change for the default balanced user.

If ASSURANCE=submission, run the pre-flight checklist below, then the verifier. The verifier's exit code is the source of truth — do NOT self-declare "audits complete" based on conversation memory.

Submission pre-flight checklist

Print this checklist verbatim at the start of Phase 6.0 and confirm each row before proceeding. This resists the common failure mode of the model skipping audits while claiming to have run them.

📋 Submission audits required before Final Report:
   [ ] 1. /proof-checker        → paper/PROOF_AUDIT.json
   [ ] 2. /paper-claim-audit    → paper/PAPER_CLAIM_AUDIT.json
   [ ] 3. /citation-audit       → paper/CITATION_AUDIT.json
   [ ] 4. bash <PRISM_REPO>/tools/verify_paper_audits.sh paper/ --assurance submission
   [ ] 5. Block Final Report iff verifier exit code != 0

<PRISM_REPO> placeholder — replace with the absolute path to your Prism clone (e.g. ~/Desktop/Prism or the path returned by dirname $(readlink ~/.claude/skills/paper-writing/SKILL.md)/../..). The path is stable across runs; store it in a shell variable if you prefer (export PRISM_REPO=~/… and use "$PRISM_REPO" in the command).

Invoking the three audits

Each sub-audit runs in a fresh Codex thread (never codex-reply, never pass prior audit output as context — this preserves reviewer independence per shared-references/reviewer-independence.md).

Each sub-audit always emits its JSON artifact, even when the content detector is negative. A detector-negative run emits verdict NOT_APPLICABLE; a silent skip is forbidden. See the "Submission artifact emission" section of each audit's SKILL.md.

Order:

1. /proof-checker "paper/" → writes paper/PROOF_AUDIT.json (emits NOT_APPLICABLE if the paper contains no theorems / lemmas / proofs)
2. /paper-claim-audit "paper/" → writes paper/PAPER_CLAIM_AUDIT.json (emits NOT_APPLICABLE if the paper has no numeric claims; emits BLOCKED if numeric claims exist but raw result files are missing)
3. /citation-audit "paper/" → writes paper/CITATION_AUDIT.json (emits NOT_APPLICABLE if no .bib file or no \cite{...} usage)

Running the verifier

bash <PRISM_REPO>/tools/verify_paper_audits.sh paper/ --assurance submission

• Exit 0 — All mandatory audits present, JSON schema-valid, hashes fresh, no blocking verdicts. Proceed to the Final Report below.
• Exit 1 — Surface paper/.prism/audit-verifier-report.json to the user verbatim, refuse to generate the Final Report, and list the specific remediation for each failing row:
  • MISSING → rerun that audit
  • STALE → paper files edited after the audit ran; rerun the affected audit
  • BLOCKING_VERDICT (FAIL / BLOCKED / ERROR) → fix the underlying issue, then rerun the audit
  • SCHEMA_INVALID → audit artifact malformed; rerun the audit

The verifier is cheap to rerun (< 1 s). After fixing any issue, rerun it before claiming green.

Optional hardening (not default)

Teams that want hook-level enforcement — i.e., the harness physically prevents a Stop event while the verifier is red — can register a Stop hook in ~/.claude/settings.json:

{
  "hooks": {
    "Stop": [
      {"command": "bash <PRISM_REPO>/tools/verify_paper_audits.sh paper/ --assurance submission"}
    ]
  }
}

This is documented here, not required. Phase 6.0's verifier-as-truth pattern is the default repo behavior.

Phase 6.1 — Final Report (runs only after the submission gate is green, or directly if assurance=draft)

# Paper Writing Pipeline Report

**Input**: [NARRATIVE_REPORT.md or topic]
**Venue**: [ICLR/NeurIPS/ICML/CVPR/ACL/AAAI/ACM/IEEE_JOURNAL/IEEE_CONF]
**Assurance**: [draft | submission]
**Submission-ready**: [yes | no]   <!-- yes iff assurance=submission AND verifier exit 0 -->
**Date**: [today]

## Pipeline Summary

| Phase | Status | Output |
|-------|--------|--------|
| 0. Assurance Setup | ✅ | paper/.prism/assurance.txt = [draft\|submission] |
| 1. Paper Plan | ✅ | PAPER_PLAN.md |
| 2. Figures | ✅ | figures/ ([N] auto + [M] manual) |
| 3. LaTeX Writing | ✅ | paper/sections/*.tex ([N] sections, [M] citations) |
| 4. Compilation | ✅ | paper/main.pdf ([X] pages) |
| 5. Improvement | ✅ | [score0]/10 → [score2]/10 |
| 4.5 Proof Audit | [PASS\|WARN\|FAIL\|NOT_APPLICABLE\|BLOCKED\|ERROR] | PROOF_AUDIT.{md,json} |
| 5.5 Paper Claim Audit | [PASS\|WARN\|FAIL\|NOT_APPLICABLE\|BLOCKED\|ERROR] | PAPER_CLAIM_AUDIT.{md,json} |
| 5.8 Citation Audit | [PASS\|WARN\|FAIL\|NOT_APPLICABLE\|BLOCKED\|ERROR] | CITATION_AUDIT.{md,json} |
| 6.0 Assurance Verifier | [OK\|STALE\|BLOCKING_VERDICT\|HAS_ISSUES\|SCHEMA_INVALID\|MISSING] per audit; exit [0\|1] overall (N/A if draft) | .prism/audit-verifier-report.json |

## Improvement Scores
| Round | Score | Key Changes |
|-------|-------|-------------|
| Round 0 | X/10 | Baseline |
| Round 1 | Y/10 | [summary] |
| Round 2 | Z/10 | [summary] |

## Deliverables
- paper/main.pdf — Final polished paper
- paper/main_round0_original.pdf — Before improvement
- paper/main_round1.pdf — After round 1
- paper/main_round2.pdf — After round 2
- paper/PAPER_IMPROVEMENT_LOG.md — Full review log
- paper/PROOF_AUDIT.{md,json} — Proof-obligation verification (always emitted at `assurance=submission`; `NOT_APPLICABLE` when no theorems)
- paper/PAPER_CLAIM_AUDIT.{md,json} — Numerical claim verification (always emitted at `assurance=submission`; `NOT_APPLICABLE` when no numeric claims; omitted in `draft` mode if Phase 5.5 detector was negative)
- paper/CITATION_AUDIT.{md,json} — Bibliography verification (always emitted at `assurance=submission`; `NOT_APPLICABLE` when no `.bib` or no `\cite{...}`; omitted in `draft` mode if Phase 5.8 detector was negative)
- paper/.prism/audit-verifier-report.json — External verifier report (submission only)

## Remaining Issues (if any)
- [items from final review that weren't addressed]

## Next Steps
- [ ] Visual inspection of PDF
- [ ] Add any missing manual figures
- [ ] Submit to [venue] via OpenReview / CMT / HotCRP

Output Protocols

Follow these shared protocols for all output files:

• Output Versioning Protocol — write timestamped file first, then copy to fixed name
• Output Manifest Protocol — log every output to MANIFEST.md
• Output Language Protocol — note: paper-writing always outputs English LaTeX for venue submission

Key Rules

• Large file handling: If the Write tool fails due to file size, immediately retry using Bash (cat << 'EOF' > file) to write in chunks. Do NOT ask the user for permission — just do it silently.
• Don't skip phases. Each phase builds on the previous one — skipping leads to errors.
• Checkpoint between phases when AUTO_PROCEED=false. Present results and wait for approval.
• Manual figures first. If the paper needs architecture diagrams or qualitative results, the user must provide them before Phase 3.
• Compilation must succeed before entering the improvement loop. Fix all errors first.
• Preserve all PDFs. The user needs round0/round1/round2 for comparison.
• Document everything. The pipeline report should be self-contained.
• Respect page limits. If the paper exceeds the venue limit, suggest specific cuts before the improvement loop.

Composing with Other Workflows

/research-lit "direction"           ← Stage 1: survey the literature
formulate research question         ← Stage 2: sharpen the thesis
develop evidence                    ← Stage 3: sources / cases / arguments
/research-review "paper topic"     ← Stage 4: iterate research
/paper-writing "RESEARCH_LOG.md"     ← Stage 5: you are here
                                         submit! 🎉

Or use /research-pipeline for the literature→review end-to-end flow,
then /paper-writing for the final writing step.

Typical Timeline

Total: ~45-90 min for a full paper from narrative report to polished PDF.