Research workflow
A CiteProof review moves through search, screening, extraction and writing. Research Mode adds three review gates, so the researcher approves the direction before the next stage starts.
Define the question and scope
Start with one answerable question and write down the boundaries before searching. The first review gate asks you to approve the problem statement and the search terms; changing them after screening is a protocol amendment, not a quiet edit.
- Population and setting. Who or what the evidence must cover.
- Intervention or exposure. The thing being studied.
- Comparator. What it is compared with, if anything.
- Outcomes. Which results matter, primary first.
- Eligibility. Study designs, date range and language limits you will apply.
The app uses these choices to build the search and to score records. They are also the criteria a reader will use to judge whether the review answered the question it set.
How the research pipeline works
Research Mode runs a 24-agent pipeline and usually takes about 15 to 25 minutes. DeepSearch runs five parallel search agents and usually takes about 5 to 10 minutes. Both are long-running jobs; you can leave the page and return to the run.
| Stage | What happens | What you get |
|---|---|---|
| 1. Search | The research question is turned into queries and sent to academic indexes from one run. | Candidate records with metadata and identifiers. |
| 2. Screen | Records are deduplicated by DOI and normalized title, then scored against the criteria. Identification, deduplication, screening and inclusion counts are tracked. | A PRISMA 2020 flow and the set of included studies. |
| 3. Extract | Findings, methods, frameworks and gaps are pulled from the screened corpus. | A structured evidence table. |
| 4. Write and check | Sections are drafted, claims are compared with source text, and citations are checked before export. | A structured report with flagged citations and evidence links. |
Human review gates
Research Mode pauses at three points. The system does not commit to downstream synthesis without your explicit approval.
| Gate | What you approve | Why it exists |
|---|---|---|
| 1. Scope and keywords | The research problem statement and the search terms, before databases are queried. | A weak query cannot be repaired by better writing later. |
| 2. Literature and framework | Synthesized findings, identified gaps and the screened evidence papers. | Catches missing strands before the paper is structured around them. |
| 3. Hypotheses | Proposed hypotheses, before methodology, analysis planning and full paper assembly. | Keeps the analysis aligned with the question you actually asked. |
DeepSearch has one checkpoint for plan approval and revisions instead of these three gates.
Sources, citations and evidence
CiteProof searches OpenAlex, Semantic Scholar, PubMed, Crossref, arXiv, CORE, Europe PMC, DOAJ, DataCite and Unpaywall. Research Mode's core academic corpus is OpenAlex, Semantic Scholar, arXiv, Europe PMC and CORE; DeepSearch uses Tavily web search across news, finance and academic results.
Identifier checks resolve DOIs against Crossref and DataCite and flag retracted literature. Citation resolution means an in-text citation matches an entry in the reference list; it does not prove that the cited paper supports the sentence. Generated claims are compared with retrieved full text by a model, and unsupported statements are flagged or rewritten. A prose lint pass flags ungrounded generalizations, vague quantifiers and process narration.
A resolved DOI can still point to a paper that does not support the claim beside it. Treat every citation check as a prompt to read the source, not as a substitute for reading it.
Reports and exports
Research Mode produces a 14-section paper, a PRISMA flow, an evidence matrix, and PDF and DOCX exports. DeepSearch produces a cited Markdown report with web sources. Exports are only available after a run reaches the export stage and your plan allows them.
If you need a separate audit record for a review you conduct in Claude Science, use CiteProof MCP. It stores the protocol, screening log, evidence rows, risk-of-bias records and claim checks, then exports one bundle.
What is automated and what needs you
| Automated | Needs a human decision |
|---|---|
| Querying sources, deduplicating by identifier and normalized title, scoring abstracts, computing PRISMA counts, formatting the report and exports. | Whether the question and eligibility criteria are sensible; whether a study should be included; whether a claim is fair; whether the final report is ready to share. |
| Checking that a DOI resolves, that a citation matches a reference, and that a quoted passage appears in supplied text. | Whether a source supports the interpretation placed on it, and whether the review's conclusions are scientifically justified. |
Limits and expectations
- Coverage. No index is complete. A search can miss relevant papers, especially non-English work, books and very recent preprints.
- Screening. Abstract relevance scores are a triage aid, not a substitute for reading the full text.
- Model judgement. Claim checks and synthesis are model-assisted and can be wrong. Always verify important statements against the source.
- Citations. A resolved citation is a metadata match. It is not evidence that the cited work supports the surrounding claim.
- Time and quota. Long runs can fail or stop at a gate. Each mode has its own monthly allowance; see plans and limits.