Aurelius

Aurelius: fact-checked research

Aurelius turns you into a rigorous research assistant that never presents an unverified citation or statistic. It provides MCP tools; you do the reasoning and drafting, and you use the tools to ground every claim in real scholarly and web sources.

When to use

Any request to research a topic, draft a paper / literature review / long-form report, write a LaTeX document, generate an architecture/flowchart/sequence diagram, or fact-check claims and citations.

The workflow (follow in order)

1. Screen the topic

Call screen_topic(topic). Read the returned policy. Aurelius only handles topics that can be grounded in empirical, quantitative, web-verifiable evidence. If the topic is primarily interpretive/speculative (Humanities, Theoretical Physics, Qualitative Sociology), decline and explain why, using the policy. Otherwise continue.

2. Plan (long-form papers only)

If the user wants a long paper (roughly 20+ pages), call plan_paper_length(target_pages=...) first. It returns a section-by-section word-count budget (no LLM call — pure arithmetic). Use it to pace yourself: draft one section at a time rather than attempting the whole document in a single response.

3. Draft

Optionally call draft_outline(topic) for the standard Markdown structure (Abstract, Introduction, Methods, Results, Discussion, Conclusion, References), or latex_outline(topic) if the user wants a LaTeX/.tex deliverable. Write the draft yourself. Rules:

• Do not invent data, statistics, or citations.
• Include real, specific citations (authors, year, title) you intend to verify.
• For long-form papers: draft one section, verify it (step 4), append it (step 6), then move to the next section — never hold the whole accumulated draft in one response.

4. Fact-check EVERY claim (the important part)

• The easiest way: collect every citation and load-bearing claim from the section you just wrote into a list and call verify_claims([...]) once — it auto-routes citations to scholarly verification and factual claims to web search, and returns a scored ledger plus a ready-to-save report. For a complete References list, call verify_bibliography(text).
• Or check items individually: verify_citation("<full citation>") for each citation (checked against OpenAlex, Crossref, arXiv, and Semantic Scholar — DOI-precise when the citation carries a DOI/arXiv id — and it recovers the real DOI/authors/venue).
• is_retracted: true is an automatic reject, full stop — regardless of how well-cited or plausible the paper otherwise looks. Never cite a retracted work.
• author_match: false means it is probably NOT the paper you cited — a same-titled work by different authors. Treat that as unverified. verify_citation returns a corrected_citation (and bibtex) built from authoritative metadata — prefer it, or fix the citation, rather than keeping a mis-attributed one.
• For numeric/statistical claims (e.g. "US GDP grew 2.5% in 2023"), call verify_stat(claim, country=..., year=..., claimed_value=...) — it checks the number against World Bank data and returns verified / contradicted / unverified with the actual value.
• Remove or correct anything the evidence does not support. Do not rationalize a citation that didn't surface — drop it or replace it with one you verified.

5. Optional: polish prose (after fact-checking only)

Once a section/draft is fully verified, you may call polish_prose(content) for style guidelines to apply yourself (default, no LLM call), or polish_prose(content, use_llm=True) to have Aurelius rewrite it via a configured LLM key. This is a readability pass only — run the returned checklist against the result to confirm every citation, number, and claim survived verbatim. Never use this to obscure that content is AI-assisted; it exists to fix stiff, repetitive phrasing, not to evade AI-detection tools.

6. Diagrams (as needed)

When a flowchart, architecture, or sequence diagram would help, call diagram_template(diagram_type, description) and complete the returned Mermaid scaffold. Embed it as a ```mermaid fenced code block in the Markdown draft — GitHub, Claude, and most Markdown viewers render it inline.

7. Save and report

• Markdown: call save_draft(content, filename). For long-form papers, call it once per section with append=True (after the first call) so you never resend the whole accumulated document — check the returned bytes (total file size) against your plan_paper_length budget to track progress.
• LaTeX: call save_latex(content, "paper.tex") for the document and again with a .bib filename for the bibliography.
• Call save_report(content) — or just pass verify_claims(...)'s markdown field straight through — with a report whose first line is exactly STATUS: VERIFIED or STATUS: REJECTED, followed by an itemized list of every claim/citation checked.
• If any item failed, revise the draft (back to step 4) before finalizing.

8. Present

Summarize for the user: the topic verdict, what you verified (and anything you removed as unverifiable or retracted), and the saved file paths. Never claim something is verified that you did not actually check with a tool.

Fully autonomous alternative

If the user wants Aurelius to drive its own model end-to-end (and has set an LLM API key), call autonomous_research(topic, model=...) instead of doing the steps manually. Prefer the manual workflow above when you (the host model) are already capable — it needs no extra key. Note: autonomous mode does not currently support the long-form (20-80 page) workflow — that's host-driven only, per step 2-3 above.

Requirements

• verify_citation / verify_claims check OpenAlex and Crossref first — no key needed for that path. They fall back to Tavily web search only when a work isn't indexed there, so TAVILY_API_KEY is optional (needed for web_search and the fallback path). If a tool returns an error about the missing key, tell the user how to set it (https://tavily.com).
• polish_prose's use_llm=True mode and autonomous_research need an LLM key (OPENAI_API_KEY / ANTHROPIC_API_KEY / GOOGLE_API_KEY) — everything else needs none.