Sci Polish

Sci-Polish — Academic Paper Polishing & De-AI

Two-stage academic paper polishing: AI trace reduction + 8-dimension quality improvement.

Hard Rules (NEVER violate)

1. NEVER modify: LaTeX commands, math environments ($...$, \[...\], equation, align, etc.), \cite{}, \ref{}, \label{}, algorithm blocks, code listings, figure/table environments (only caption text is editable)
2. NEVER fabricate: data, citations, experimental results, baselines, metrics, author names
3. NEVER remove: author's technical claims without explicit permission
4. NEVER introduce: Tier 1 AI words (see references/word-choice-anti-ai.md) during Phase B edits
5. Mark as [PENDING VERIFICATION]: any claim you cannot verify from the text
6. Preserve meaning: the exact semantic content of every technical statement must be unchanged
7. Output format: revised text + polish report. No preamble or trailing changelog unless user asks
8. When uncertain about domain terminology: keep the original, flag for user review

Execution Modes

Step 0: Input Processing

1. Identify input format:
   • PDF → extract visible text (preserve section structure)
   • LaTeX → extract visible paragraphs (protect all commands)
   • Markdown / plain text → use directly
2. Identify protection zones: math environments, citations, labels, algorithm blocks, code blocks, figure/table environments
3. Ask target journal (optional): affects Dimension 8 scoring. If not specified, use general academic standards.
4. Confirm execution mode: full / deai-only / polish-only / polish-select

Step 1: Phase A — De-AI Editing

Reference files:

• references/ai-detection-patterns.md — pattern library for scanning
• references/de-ai-strategies.md — rewriting strategies
• references/word-choice-anti-ai.md — word replacement table

1.1 Scan

Scan text against ai-detection-patterns.md. Categorize findings:

• High priority: Tier 1 words, template phrases, burstiness deficit (SD < 8), RLHF fingerprints
• Medium priority: Tier 2 threshold violations, structural symmetry, transition stacking
• Low priority: Tier 3 context-dependent, punctuation fingerprints, minor style uniformity

1.2 Rewrite (Priority Order)

Follow the strategy priority from de-ai-strategies.md:

a. Burstiness enforcement — restructure sentence lengths (Strategy 1)
   Target: SD > 8, range ≥ 20, mid-band < 50%

b. Vocabulary de-AI — replace Tier 1 words, compress template phrases (Strategy 2)
   Consult word-choice-anti-ai.md for replacements

c. Sentence restructuring — vary syntax patterns (Strategy 3)
   No 3+ consecutive same-structure sentences

d. Concept concretization — add specifics where available (Strategy 4)
   Replace vague claims with data from the paper itself

e. Argumentation enrichment — add depth to claims (Strategy 5)
   Insert boundary conditions, alternative interpretations

f. Perplexity elevation — introduce unpredictability (Strategy 6)
   Unconventional openers, domain-specific minor terms

g. Style disruption — break uniformity (Strategy 7)
   Vary paragraph density and register between sections

1.3 Self-Audit

After rewriting, re-scan output against ai-detection-patterns.md:

• Any remaining High-priority patterns? → fix them
• Burstiness check: SD > 8? Range ≥ 20? Mid-band < 50%?
• Protection zones intact? (LaTeX commands, math, citations unchanged)
• Meaning preserved? (no semantic drift from original)

Step 2: Phase B — 8-Dimension Polish

Reference file: references/polish-dimensions.md

2.1 Score All Dimensions (1-5)

Rate each dimension based on the checkpoints in polish-dimensions.md:

1. Grammar & Syntax
2. Academic Tone
3. Coherence & Flow
4. Conciseness
5. Terminology
6. Structure & Formatting
7. Clarity of Argument
8. Journal Compliance

2.2 Edit Dimensions Scoring ≤ 3

Edit in this order (foundation → logic → style → format):

Dim 1 (Grammar) → Dim 7 (Clarity) → Dim 3 (Coherence) →
Dim 4 (Conciseness) → Dim 2 (Tone) → Dim 5 (Terminology) →
Dim 6 (Structure) → Dim 8 (Journal)

2.3 Preservation Check

During Phase B editing, verify Phase A improvements are maintained:

• Burstiness NOT reduced (sentence length variety preserved)
• No Tier 1 AI words reintroduced
• No template phrases reintroduced
• Punctuation fingerprints still clean (em dash count, etc.)

2.4 Re-Score

Rate edited dimensions again. Record "Before" and "After" scores for the report.

Step 3: Output

3.1 Revised Text

Output the complete polished text with all modifications applied. Format matches the input format (LaTeX → LaTeX, Markdown → Markdown, etc.).