agent-module-academic-research-pipeline

name	agent-module-academic-research-pipeline
description	Module: Autonomous Academic Research and Paper Production Pipeline

Module: Autonomous Academic Research and Paper Production Pipeline

1. Module Metadata

• Module ID: agent.module.academic-research-pipeline
• Version: 1.0.0
• Maturity: production
• Scope: End-to-end autonomous pipeline for literature search, structural planning, writing, peer-review simulation, citation management, and LaTeX compilation for academic papers.
• Primary outcomes:
  • Human-quality, production-ready LaTeX academic manuscripts.
  • Zero presence of AI writing patterns, hallmarks, or repetitive structures.
  • Verified and cross-checked citations with automated BibTeX management.
  • Architecturally sound paper structuring via a systematic 5-phase pipeline.

2. Mission and Applicability

Use this module to orchestrate an AI agent codebase for writing and researching high-quality academic papers (Q1 journal and conference level).

Apply when:

• Synthesizing broad academic literature (Semantic Scholar, arXiv, CrossRef).
• Drafting technical sections, related works, and methodologies.
• Reviewing, critiquing, and humanizing prose to meet academic standards.
• Managing LaTeX formatting, building, and compilation workflows.

Do not apply directly when:

• Writing non-academic content (e.g., blog posts, generic documentation).
• Performing raw code generation lacking an academic or theoretical context.
• Generating creative prose without evidence-based constraints.

3. Architecture Pattern

• Pattern: 5-Phase Autonomous Pipeline + Skill Auto-Trigger Orchestration
• Core design rules:
  • Decoupled Execution: Research, Structuring, Writing, Quality Assurance, and Compilation are isolated phases.
  • Continuous Humanization: Anti-AI pattern scanning ("humanizer") runs immediately after every drafted section, not just at the end.
  • Strict Constraints: Enforcement of mandatory writing rules (no colons/semicolons in prose, academic humility, full formula explanations).
  • Parallelization: Independent sections (e.g., related work and methodology) can be written or reviewed simultaneously.

4. Implementation Workflow

Phase A: RESEARCH

1. Broad Search: Query academic databases (Semantic Scholar, arXiv, CrossRef) for the topic.
2. Filter & Categorize: Sort results by relevance, recency, and citation count. Group into thematic buckets.
3. Gap Analysis: Correlate what has been studied versus what is missing. Identify the novel contribution.
4. Output: Generate research_notes.md containing categorized papers, gaps, and novelty assessments.

Exit criteria:

• Literature review yields >50 high-relevance papers with explicit novelty checks.

Phase B: STRUCTURE

1. Outline Generation: Create a section-by-section outline based on research_notes.md.
2. Dependency Mapping: Determine logical writing order (e.g., Intro -> Methodology -> Results -> Related Work -> Discussion -> Conclusion -> Abstract).
3. Placements: Identify exact locations for figures and tables (must be near first reference).
4. Output: Complete structured outline with quality checkpoints inserted.

Exit criteria:

• Every section has a defined purpose, key arguments, and mapped citations/evidence.

Phase C: WRITING

1. Section Drafting: Follow the outline to draft sections concurrently (up to 4 independent sections at once).
2. Humanization Pass: Immediately run anti-AI pattern analysis on the drafted section and rewrite flagged segments.
3. Citation Insertion: Automatically inject \cite{} references and validate against the references.bib file.
4. Transition Linking: Generate seamless textual transitions between fully drafted sections.

Exit criteria:

• Complete main.tex with all text, citations, and figure/table references integrated.

Phase D: QUALITY

1. Devil's Advocate Review: Scan with a 13-dimension academic critique. Classify issues as Critical, Important, or Minor.
2. Global AI Pattern Scan: Run a secondary pass across the entire document for 25+ categories of AI writing signatures.
3. Simulated Peer Review: Execute multi-persona (methodological, theoretical, practical) reviews outputting scores and insights.
4. Resolution Cycle: Fix Critical and Important issues, triggering localized humanizer passes on structural changes.

Exit criteria:

• Document passes all reviewer personas and citation validation scripts without Critical/Important flags.

Phase E: COMPILE AND SUBMIT

1. Compilation: Execute LaTeX build chain (e.g., pdflatex -> bibtex -> pdflatex).
2. Format Verification: Ensure constraints (page count, font size, margin limits) for the target venue (e.g., IEEEtran, NeurIPS) are met.
3. Error Recovery: Automatically parse and fix LaTeX build errors or missing references.
4. Output: A clean, submission-ready main.pdf.

Exit criteria:

• Clean PDF generation with no outstanding warnings or formatting violations.

5. Decision Framework

Decision Area	Preferred Option	Alternative	Selection Rule
AI Pattern Detection	Per-section humanizer pass	End-of-pipeline single pass	Per-section prevents AI hallmarks from compounding and polluting subsequent context.
Writing Sequence	Intro -> Methods -> Results -> Related Work	Linear (Intro -> Related -> ...)	Writing 'Related Work' later ensures better positioning against the paper's actual contribution.
Citation Validation	Automated BibTeX script validation	Manual LLM cross-checking	Scripts prevent hallucinated citations and ensure strict LaTeX compilation compatibility.
Prose Constraints	Strict rule enforcement (No colons, no generic transition words)	Default LLM generation	Enforcing strict stylistic rules is the only reliable way to pass human scrutiny.

6. Validation Strategy

Functional Validation

• Verify that every technical term is explained in plain English prior to its first use.
• Verify that every LaTeX formula includes a comprehensive "where" clause defining all symbols.
• Ensure all figures and tables appear on the same page or immediately following their first text reference.

Failure Validation

• Check behavior when LaTeX compilation throws an error (must enter auto-fix loop).
• Verify graceful degradation if an academic database (e.g., CrossRef) rate limits the search.

Contract/Security/Performance Validation

• Validate that the AI writing vocabulary strictly avoids the 25+ banned words (e.g., landscape, vibrant, delve, nuanced, tapestries).
• Confirm zero usage of em-dashes, colons, or semicolons in standard prose.

7. Benchmarks and SLO Targets

• AI Detection Score: < 5% false-positive probability on standard AI-detectors.
• Citation Integrity: 100% of \cite{} commands resolve successfully in references.bib.
• Prose Burstiness: Sentence length variance must be high; no >4 consecutive sentences of identical length.
• Compile Success Rate: 100% successful PDF generation for final output.

8. Risks and Controls

• Risk: Hallucinated citations or unverified claims.
  • Control: Integrate citation-management scripts and novelty-assessment cross-referencing.
• Risk: "Sterile" or overly-robotic academic tone.
  • Control: Enforce narrative flow rules (Story Flow rule) and utilize creative-prose critique skills to ensure readability across expertise levels.
• Risk: Scope creep causing disjointed storylines between intro and conclusion.
  • Control: Utilize the dependency mapping created generated during the STRUCTURE phase.

9. Agent Execution Checklist

• Determine paper metadata (Title, Authors, Target Venue, Format).
• Initialize master configuration mapping (Skill Auto-Trigger Map).
• Execute Phase 1 (RESEARCH): Generate research_notes.md.
• Execute Phase 2 (STRUCTURE): Build dependent outline with quality checkpoints.
• Execute Phase 3 (WRITING): Draft sections, apply per-section humanizers, validate \cite{}.
• Execute Phase 4 (QUALITY): Run devil's advocate review, simulated peer review, and global AI scan.
• Execute Phase 5 (COMPILE): Build LaTeX, format check, and output PDF.

10. Reuse Notes

Adapt only:

• Target venue templates (e.g., swapping IEEEtran for a generic two-column conference format).
• Compilation commands based on local environment (Local LaTeX vs. Dockerized compilation).
• The stringency of the devil's advocate review depending on the target publication tier (Q1 journal vs. Workshop paper).