Sisyphus Academica — The Research Paper Writing Army
20+ specialized agents. 6 novelty engines. 10 adversarial reviewers. Zero hallucinated citations. Zero AI-isms.
⭐ If you write research papers, star this repo — it will save you weeks of work.
Not a writing assistant. Not a chatbot with a LaTeX plugin. A self-coordinating swarm of 20+ specialized agents that produces publication-ready research papers with genuine novelty, zero hallucinated citations, and no detectable AI-written patterns.
git clone https://github.com/argahv/sisyphus-academica.git && cd sisyphus-academica
bash install.sh
# Select "research-director" → "write a paper about [topic]"
CLI Tools (No Agent Required)
The Python CLI works standalone — no OpenCode or agent platform needed, and no dependencies beyond the Python 3.10+ standard library:
git clone https://github.com/argahv/sisyphus-academica.git && cd sisyphus-academica
pip install -e . # or run bin/sisyphus without installing
sisyphus demo # Run all five quality gates on a bundled example paper (offline, no API keys)
sisyphus search QUERY # Search arXiv, Semantic Scholar, CrossRef and OpenAlex in parallel
sisyphus verify FILE # Verify citations in a paper JSON, .bib or .tex against 2+ APIs
sisyphus bibtex DOI # Generate BibTeX from a DOI
sisyphus configure # Set up API keys in .env
The same CLI runs the paper pipeline for the agents, and you can call it yourself:
sisyphus new "TOPIC" # create out/papers/<slug>/ for one paper
sisyphus gate patterns draft.tex # 41 AI-writing patterns + em dash count
sisyphus gate statistics paper.tex # recompute p-values, check CIs and abstract numbers
sisyphus gate all RUN # all five gates on a paper run
sisyphus status RUN # which stages and gates are done
sisyphus build RUN # LaTeX + BibTeX + PDF, refused until every gate passes
Portable Agent Skills (Works with Any Agent)
The novelty engines and reviewer personas are packaged as standalone agent skills — drop them into any agent that reads SKILL.md (Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and more):
git clone https://github.com/argahv/sisyphus-academica.git
# Manual: copy individual skill directories
cp -r skills/contrarian ~/.claude/skills/
# Or use npx (see below)
Or install any skill with a single command:
npx skills add argahv/sisyphus-academica # Browse and select interactively
npx skills add argahv/sisyphus-academica -s '*' # Install all 17 skills
npx skills add argahv/sisyphus-academica -s contrarian # Install one skill
Then invoke directly in your agent:
/contrarian "The claim: 'Attention is all you need'"
/cross-pollinator "Problem: How to reduce LLM hallucination"
/heretic "Paper: 'Scaling Laws for Neural Language Models'"
6 novelty engines + 10 reviewers + academic humanizer = 17 portable skills.
💡 Prefer a standalone repo? The 6 novelty engines and 5 reviewer personas are also available as a dedicated skill pack at argahv/novelty-skills — install via npx skills add argahv/novelty-skills.
For Humans
Strongly recommended: let an LLM agent install this for you. The setup involves API key configuration, agent deployment across 25 specialized agents, voice profile calibration, and provider selection — humans fat-finger these. An LLM agent reads the full guide and walks every step correctly.
The agent will greet you with oMoMoMoMo... and ask one question at a time:
- Which platform? (OpenCode recommended)
- Do you have oh-my-openagent?
- Claude subscription? (Pro / Max 20x / None)
- OpenAI subscription? (Plus / None)
- Semantic Scholar API key? (free — boosts rate limits)
- LaTeX for PDF? (or Docker / skip for .tex only)
- Writing sample for voice calibration? (optional)
Then it installs, configures, and verifies everything automatically.
Paste this prompt into Claude Code, AmpCode, Cursor, or any agent:
Install and configure Sisyphus Academica by following the instructions here:
https://raw.githubusercontent.com/argahv/sisyphus-academica/main/docs/guide/installation.md
Manual install — one line, no agent needed
If you prefer to run the installer yourself:
git clone https://github.com/argahv/sisyphus-academica.git && cd sisyphus-academica
bash install.sh
The installer will ask you the same 6 questions and configure everything automatically. Run with --yes to skip all prompts and use defaults:
bash install.sh --yes
The Numbers (Not Claims)
| Metric | GPT-4 / NotebookLM | Sisyphus Academica |
|---|
| Papers surveyed per run | 10-50 | 500+ (5 parallel scouts) |
| Citation hallucination rate | ~30-40% | 0% (verified against 2+ APIs) |
| Novelty generation | None ("what's the gap?") | 6 engines × 50+ hypotheses |
| Adversarial review | None | 10 distinct personas |
| AI-text detection | Post-hoc (chatgpt.com paste) | Generation-time (41 patterns) |
| Voice calibration | None | Learns from your writing samples |
| Output format | Raw text / Google Doc | LaTeX PDF with verified BibTeX |
| Time to first draft | 5-20 min | 30 min - 4 hours (reviewed) |
Pipeline stats from our SIREN paper (view output):
100+ papers surveyed → 6 novelty engines → 50+ hypotheses → 10 adversarial reviewers → 4 revision rounds → 13-page PDF with 20 cited references and 2 embedded figures. The example predates the code-enforced gates; its README lists what the current gates flag in it.
Quick Start
# 1. Clone + install
git clone https://github.com/argahv/sisyphus-academica.git
cd sisyphus-academica
bash install.sh
# 2. Configure API keys (free, optional; install.sh asks for them too)
sisyphus configure --semantic-scholar-key KEY --crossref-email [email protected]
# 3. Write a paper
# OpenCode → agent tab → select "research-director"
# → type: "write a paper about transformer efficiency"
Prerequisites: OpenCode (or compatible agent platform), Python 3.10+, LaTeX (optional, for PDF).
Provider-agnostic — works with any OpenAI-compatible or Anthropic API. Edit config/agent-config.json to switch models, then re-run bash install.sh to apply them to the installed agents:
{
"default_model": "anthropic/claude-sonnet-4",
"agents": {
"research-director": {
"model": "anthropic/claude-opus-4",
"fallback_models": [{"model": "anthropic/claude-sonnet-4"}]
},
"literature-scout": { "model": "openai/gpt-4o" }
}
}
Agents without their own model use default_model. Set LLM_PROVIDER in .env to give every agent the same model.
Architecture
┌──────────────────────────────┐
│ Research Director │
│ (orchestrator) │
└────────┬─────────────────────┘
│
┌────────────────────┼─────────────────────────┐
▼ ▼ ▼
┌───────────┐ ┌────────────────┐ ┌──────────────────────┐
│ Literature│ │ 6 Novelty │ │ Gap Analyzer │
│ Scout ×5 │ │ Engines │ │ + Methodology │
└───────────┘ └────────────────┘ └──────────────────────┘
│
┌────────────────────┼─────────────────────────┐
▼ ▼ ▼
┌─────────────────────────────────────────────────────────────┐
│ Parallel Writing Swarm (5 agents + 41 Humanizer patterns) │
└──────────────────────────┬──────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────┐
│ Verifier: Citations × Statistics × AI-Pattern Detection │
└──────────────────────────┬──────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────┐
│ Adversarial Review: 10 reviewer personas (parallel) │
└──────────────────────────┬──────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────┐
│ Style Auditor: Humanizer certification, em dash zero check │
└──────────────────────────┬──────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────┐
│ Formatter: LaTeX template, BibTeX, figures, PDF │
└─────────────────────────────────────────────────────────────┘
The Novelty Engines (The Moat)
Six engines that think like no human can:
- The Contrarian — Inverts every well-established claim in the field. Generates 10 counter-hypotheses.
- The Cross-Pollinator — Imports solutions from 15 distant fields (astrodynamics, epidemiology, music theory, immunology, 15th-century shipbuilding...).
- The Assumption Excavator — Finds unstated assumptions and tests what breaks if they're false.
- The Counterfactual Generator — Rewrites the field's history without the most-cited papers.
- The Paradox Sifter — Cross-references every "Limitations" section to find ignored contradictions.
- The Heretic — Crown jewel. Generates 50 wild hypotheses from title+abstract alone, scores each against the actual paper, and finds the "haunting idea" — what the paper should have been.
The Adversarial Reviewers
Each paper is independently reviewed by 10 distinct personas running in parallel. All 10 must pass before formatting.
| Persona | Focus | Typical Critique |
|---|
| Theorist | Formal proofs, mathematical rigor | "Where's the formal proof?" |
| Empiricist | Experimental design, baselines | "Your baseline is wrong" |
| Pragmatist | Practical applicability | "Does this matter in practice?" |
| Skeptic | Default: results are wrong | "Show me error bars" |
| Historian | Prior art, citation accuracy | "This was done in 1972" |
| Methodologist | Statistical methodology | "Your test assumes normality" |
| Ethicist | Societal implications | "What are the downsides?" |
| Competitor | Novelty relative to existing work | "Minor mod of our 2023 paper" |
| Student | Clarity and accessibility | "I don't understand section 3" |
| Dreamer | "What if you went further?" | "You stopped too early" |
Quality Gates
Every paper passes through 5 hard gates. They are enforced in code (sisyphus gate), and sisyphus build refuses to format a draft unless all five passed on that exact draft. If any gate fails, the paper goes back to revision.
- Citation Verification — Every reference looked up in Semantic Scholar, CrossRef, OpenAlex and arXiv. Title, year and authors must match in 2+ sources. Cited keys without a verified entry block the paper.
- Statistical Audit — p-values recomputed from t, F, χ², z and r statistics; "significant" with p > .05, impossible values, estimates outside their CI, missing multiple-comparison correction and abstract numbers missing from the body are all errors. Every reported test must appear in the analysis output.
- AI-Pattern Detection — 41 Humanizer patterns scanned. Density must be < 2 violations per 1000 words, and zero em dashes.
- Style Audit — Zero em dashes. Pattern density < 1/2000 words. Voice must match the author's writing profile (when one exists).
- Adversarial Review — All 10 reviewer personas must recommend acceptance with no open must-fix items. Not a subset. All 10.
Live Example: SIREN Paper
The pipeline was run to produce a full paper on Intent-Based Blockchain Execution via Agentic RAG and Swarm Consensus. Complete output in examples/siren-paper/:
| File | Description |
|---|
siren-paper.pdf | 13-page compiled paper |
siren-paper.tex | LaTeX source (508 lines, 20 cited references) |
figures/*.pdf | 3 figures (2 embedded in the paper) |
README.md | Pipeline summary with review scores |
Review scores progressed from Avg 4.6/10 → 8/10 across 4 revision rounds.
FAQ
Q: Does this require a specific LLM provider?
No. Edit config/agent-config.json to use any OpenAI-compatible or Anthropic API.
Q: Can I add my own LaTeX template?
Yes. Add templates/<venue>/template.tex (plus any .sty, .cls or .bst files). Put <<title>>, <<authors>> and one <<section_name>> placeholder per section where the content goes; sisyphus new "topic" --venue <venue> then plans those sections and sisyphus build fills them.
Q: How long does a paper take?
30 minutes to 4 hours depending on LLM speed, literature volume, and revision rounds.
Q: The output sounds too AI-like. What do I do?
Put 2-3 paragraphs of your published writing in data/voice-profile/*.txt and run sisyphus voice. Writers get the measured profile (sentence length, punctuation, hedging, use of "we") as guidance, and the style gate checks the draft against it.
Q: Can I use this without OpenCode?
The agents are OpenCode-compatible, but the sisyphus CLI (search, citation verification, BibTeX, the five quality gates and the LaTeX build) works standalone.
Q: How do I contribute?
See CONTRIBUTING.md. Good first issues are tagged. Template stubs need filling, the PyPI package needs publishing, and more reviewer personas are welcome.
Directory Structure
sisyphus-academica/
├── orchestrator/ # Research Director agent (the conductor)
├── subagents/ # Core writing pipeline agents (writer, verifier, etc.)
├── novelty-engines/ # 6 novelty generation agents
├── reviewers/ # 10 adversarial reviewer personas
├── skills/ # 17 portable skill files (also available as standalone [novelty-skills](https://github.com/argahv/novelty-skills) repo)
├── tools/ # Python toolchain (stdlib only)
│ ├── literature_client.py # Multi-source lit search
│ ├── citation_verifier.py # Citation verification + BibTeX
│ ├── patterns.py # 41 AI-writing patterns + em dashes
│ ├── stats_audit.py # Statistical audit
│ ├── voice.py # Voice profile build + check
│ ├── reviews.py # 10-reviewer gate
│ ├── gates.py / build.py # Gate runner, LaTeX/PDF formatter
│ └── workspace.py # Paper run layout + stage handoff checks
├── src/sisyphus/ # `sisyphus` CLI + bundled demo paper
├── bin/sisyphus # Run the CLI without installing
├── templates/ # LaTeX venue templates (add yours)
├── config/ # Agent configuration
├── examples/siren-paper/ # Full pipeline output (13-page paper)
├── data/ # Your voice samples + profile (created by configure, gitignored)
├── tests/ # Python unit tests
├── docs/ # GitHub Pages documentation
├── docker-compose.yml # LaTeX + dev environments
└── pyproject.toml # Package metadata
Development
pip install -e ".[dev]"
python -m pytest tests/ -v # offline: all API calls are stubbed
flake8 tools/ src/ tests/ --max-line-length=100
# PDF without a local TeX install
sisyphus build out/papers/<slug> --docker
Acknowledgments
- Humanizer by blader — the 30-pattern AI-detection skill this system builds on
- Novelty Skills — standalone thinking tools for AI agents (separate repo)
- OpenCode + OhMyOpenAgent — agent orchestration platform
- All six novelty engines were inspired by cognitive diversity research
License
MIT — see LICENSE for details.
⭐ Star this repo if you write research papers — it helps others find it.
