Academic Writing Pipeline#

End-to-end academic paper production with built-in quality gates and multi-agent review.

Pipeline Overview#

text

Research → Write → Integrity Check → Review → Revise → Summary

Each stage has defined inputs/outputs and quality gates. The pipeline is non-linear: stages may loop (Review → Revise → Re-Review) until quality threshold met.

Stage Details#

Stage 1: Research#

Goal: Gather and organize evidence.

Actions:

1. Literature search via Semantic Scholar API
2. Filter by relevance score ≥ 0.5
3. Grade evidence level (A: meta-analysis, B: RCT, C: observational, D: opinion)
4. Output: research/evidence.json

Script: scripts/research.py

Stage 2: Write#

Goal: Generate structured manuscript.

Actions:

1. Build argument chain from evidence
2. Generate sections: Abstract, Introduction, Methods, Results, Discussion
3. Track citation markers for each claim
4. Output: draft/manuscript.md

Script: scripts/write.py

Stage 3: Integrity Check (CRITICAL)#

Goal: Anti-hallucination verification.

Check types:

• Citation verification: DOI exists? Authors match? Year correct?
• Data verification: Numbers match tables/figures?
• Claim verification: Evidence supports assertion?

Threshold: Must pass 100% of checks to proceed.

Script: scripts/integrity_check.py

APIs used:

• Semantic Scholar ( https://api.semanticscholar.org)
• CrossRef DOI ( https://api.crossref.org/works/)

Stage 4: Review (5-Person Panel)#

Agents:

RoleFocusScore WeightEditor-in-ChiefContribution, journal fit30%MethodologyMethods, stats, reproducibility25%Domain ExpertRelated work, theory20%Devil's AdvocateStrongest counter-arguments15%SynthesizerMerge opinions, roadmap10%

Decision mapping:

• ≥80: Accept
• 65-79: Minor Revision
• 50-64: Major Revision
• <50: Reject

Script: scripts/review.py

Stage 5: Revise#

Goal: Address reviewer feedback.

Actions:

1. Parse Synthesizer roadmap
2. Generate revision plan with priorities
3. Rewrite affected sections
4. Re-run Integrity Check

Script: scripts/revise.py

Stage 6: Process Summary#

Goal: Auditable record.

Output: summary.json containing:

• Timeline of each stage
• Decision points and scores
• Integrity check results
• Reviewer scores and comments
• Revision history

Script: scripts/summary.py

Configuration#

Edit config.yaml for:

• Model selection (default: qwen3.5-plus)
• Temperature (default: 0.3 for stability)
• Review thresholds
• API keys

Usage#

bash

# Full pipeline
python scripts/main.py --topic "your research topic"

# Single stage
python scripts/main.py --stage integrity-check --input draft/manuscript.md

# With custom config
python scripts/main.py --config custom_config.yaml

Quality Gates#

GateRequirementAction on FailEvidence≥5 Grade A/B sourcesReturn to ResearchIntegrity100% verificationReturn to WriteReview≥65 scoreLoop ReviseFinal Integrity100% verificationBlock submission

Key Principles#

1. Integrity First: Citation verification is non-negotiable
2. Quantified Review: Scores enable objective decisions
3. Loopable Pipeline: Revision cycles until threshold met
4. Auditable Output: Process Summary for journal submission

Reference Files#

• references/review_rubric.md - Detailed scoring criteria
• references/evidence_levels.md - Evidence grading standards
• references/citation_styles.md - Journal formatting guides