# studydiff [Health: Active]

**Category:** 🔬 Research  
**Repository:** https://github.com/nickjlamb/studydiff  
**GitHub Stars:** 1  
**npm Downloads (last month):** 155  
**Views:** 1  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/studydiff

## Description
Explain why two scientific papers disagree, every claim grounded in a verbatim source quote.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "studydiff": {
    "command": "npx",
    "args": ["-y","studydiff-mcp"]
  }
}
```

## Documentation

## What the studydiff MCP server does

The studydiff MCP server examines two scientific studies that appear to disagree. It extracts details of each study’s design, compares those details, and presents the dimensions where the papers diverge. Examples described by the project include cell type, dose, follow-up period, assay, and analysis choices.

Each reported value is accompanied by the sentence from the source that supports it. If a paper does not state a field, the result marks it as not reported rather than filling the gap with an inference. Identical dimensions are also useful: they can help rule out some methodological explanations for the disagreement.

The intended user is a bench scientist deciding which of two conflicting papers to examine more closely before designing an experiment. The project also provides a live demo and an npm package named `studydiff-mcp`.

## How it works

The comparison starts with two studies. StudyDiff extracts design information from each source, identifies divergent dimensions, and associates extracted values with verbatim evidence. A deterministic grounding check then verifies that extracted values and explanations are supported by quoted source text and traceable numbers. Statements that fail this check are downgraded before they can be used as explanations.

The current behavior deliberately avoids naming one primary driver of the disagreement. An earlier ranking approach used a fixed priority that placed assay differences first. Evaluation showed that this ranking selected assay in a way indistinguishable from repeatedly guessing assay, including on cases where the established cause was something else. The application therefore presents divergent dimensions as an unranked list. Selecting the most important cause still requires domain knowledge outside the tool.

## Tools and capabilities

The documented capabilities are:

- Compare the designs of two scientific studies.
- Extract study details and expose differences between the papers.
- Show verbatim source sentences for extracted values and explanations.
- Mark information absent from the source as not reported.
- Use identical dimensions to help rule out possible causes.
- Apply deterministic grounding checks to extracted claims.

The project’s benchmark separates extraction from explanation ranking. On both the development and held-out sets described in the README, the prior ranking identified none of 25 non-assay established causes. The extracted candidate list nevertheless contained the established cause in 10 of 15 development cases and 11 of 15 held-out cases, according to the reported oracle ceilings.

## Limitations and notes

The studydiff MCP server does not reliably determine which difference caused a disagreement. Its published evaluation reports strict top-1 accuracy of 13.3% on both a 15-case development set and a separate 15-case held-out set.

Those figures measure the former ranked-driver behavior, not the current unranked presentation. They are important when deciding how much to rely on the output: use the results to inspect methodological differences and their supporting passages, not as an authoritative causal verdict.

The README includes commands for reproducing the benchmark offline and for checking benchmark integrity, including `npm run eval` and `npm run eval:selftest`.

_Full upstream README: https://allmcps.com/mcp/studydiff/readme_