Lunheng-quick — single-reviewer 2-round paper polish. Use when paper is structurally OK and only needs writing/clarity polish. Lighter than full lunheng (no architect, no writer/refiner split, no 8-evaluator bench). Use when user says "quick polish", "快速润色", "lunheng quick", or wants a fast review-fix-recompile loop without the full multi-agent setup.

Lunheng-Quick — Single-Reviewer Polish Loop

A lightweight version of lunheng for papers that don't need full restructuring.

Operate on the paper at: $ARGUMENTS

When to Use Each

Paper state	Use
Major structural issues (sections out of order, figure–text mismatch, terminology drift)	`lunheng`
Polished structurally, just needs writing/clarity sweep	`lunheng-quick` ← (this)
Want anchored 8-dim scoring	`lunheng`
Want a fast yes/no readiness check	`lunheng-quick`

Constants

MAX_ROUNDS = 2
STOP_THRESHOLD = 7.0 (composite over the same 8 dimensions)
WORKSPACE = <paper_dir>/lunheng_quick_workspace/

Workflow

Step 0: Snapshot

mkdir -p <paper_dir>/lunheng_quick_workspace
cp <paper_dir>/main.pdf <paper_dir>/lunheng_quick_workspace/main_round0.pdf 2>/dev/null

Step 1: Round 1 — Single Reviewer

Spawn ONE Claude sub-agent (subagent_type: general-purpose) with this prompt:

You are a senior reviewer for a Q1 journal in this paper's field.
Score the paper at <paper_dir>/<main.tex> on the 8-dimension Lunheng rubric.

Anchors (1-10 scale, see docs/EVALUATION_RUBRIC.md):
D1 Soundness | D2 Significance & Originality | D3 Clarity & Organization
D4 Experimental Substance | D5 Reproducibility | D6 Citation Quality
D7 Visual & Tabular Communication | D8 Ethics, Limitations & Broader Impact

For each dimension, return:
- score (1-10)
- 1-2 most critical issues
- 1-2 specific actionable fixes
- score_change_criteria ("Score will rise to X if Y")

Compute composite R = mean(D1..D8).
Provide an Overall verdict (NeurIPS-style 1-6 scale): 6=top-tier, 5=strong accept,
4=borderline accept, 3=borderline reject, 2=reject, 1=strong reject.

Save the full review to lunheng_quick_workspace/round1_review.md.

Step 2: Implement Round 1 Fixes

Apply CRITICAL fixes first, then MAJOR. Common fix patterns:

Issue	Fix
Data dumping	Rewrite paragraphs as claim → evidence → interpretation
Overclaim	Soften language ("证明" → "表明", "validate" → "demonstrate")
Terminology drift	Choose canonical form, replace globally
Missing tool citations	Add to .bib, cite at first mention
Dangling \ref or \cite	Remove or define the missing target
Caption only labels axes	Rewrite with self-contained description

Step 3: Recompile

xelatex -interaction=nonstopmode <main>.tex
bibtex <main>
xelatex -interaction=nonstopmode <main>.tex
xelatex -interaction=nonstopmode <main>.tex

Verify 0 undefined refs/citations.

Step 4: Round 2 — Fresh Reviewer

Spawn a fresh sub-agent with no memory of Round 1. Use the same prompt as Step 1, on the revised paper.

Compute $R_2$. If $R_2 \geq 7.0$: STOP.

Step 5: Document

Write <paper_dir>/LUNHENG_QUICK_LOG.md:

# Lunheng Quick Polish Log

## Composite Scores

| Dimension | R0 | R1 | R2 |
|-----------|-----|-----|-----|
| D1 Soundness | a | b | c |
| ... |
| **R (mean)** | **R0** | **R1** | **R2** |

## Round 1 Findings & Fixes
<critical issues + what was changed>

## Round 2 Findings (final)
<remaining issues + verdict>

Key Differences from Full Lunheng

Feature	lunheng	lunheng-quick
Architect agent	Yes	No
Writer/Refiner split	Yes	No (just edits)
Evaluator count per round	8 (parallel)	1
Visual contract	Persistent JSON	Implicit
Sub-agent calls per run	10–25 per round	1 per round
Best for	Deep restructuring	Polish pass

lunheng-quick