Skip to content
zaksblingPublic

About

Regulatory feedback for real-world evidence, from public FDA reviews and EMA assessment reports.

Resources

Stars

0 stars

Watchers

0 watching

Forks

 
 

Repository files navigation

Reg4RWE

Public FDA reviews and EMA assessment reports, read as regulatory feedback on real-world evidence. The navigator lets you move through how the evidence was used, the question it was asked to answer, the setting, and what the regulator accepted or criticised.

Some records still need a line-by-line check. Those are marked Needs manual review. FDA’s own published example tables are included and compared with this corpus.

Open the app

The hosted app is at https://www.reg4rwe.com. reg4rwe.com redirects there.

To run it locally:

python3 serve.py

Then open http://127.0.0.1:8765.

Appraise your own design

The skill reg4rwe holds a proposed study against this feedback. It returns the closest precedents, the regulator’s words where a passage was extracted, and the feedback that recurs across those precedents. It then asks the assistant to appraise the design from those points only.

Activate it by copying skills/reg4rwe into the skills folder your client reads:

Client Folder
Claude Code .claude/skills/reg4rwe or ~/.claude/skills/reg4rwe
Cursor .agents/skills/reg4rwe in this repository, or ~/.cursor/skills/reg4rwe
GitHub Copilot .github/skills/reg4rwe
Grok the same skill file, plus the MCP server below if the client accepts MCP

Copies for Claude Code, Cursor, and Copilot are already in this repository (.claude/skills, .agents/skills, and .github/skills).

MCP server

The server speaks the Model Context Protocol on standard input and output. The tool is appraise_rwe_design. Point the client at this checkout:

{
  "mcpServers": {
    "reg4rwe": {
      "command": "python3",
      "args": ["mcp/server.py"],
      "cwd": "/absolute/path/to/Reg4RWE"
    }
  }
}

The same block works in Claude Desktop (claude_desktop_config.json), Claude Code, Cursor, and VS Code Copilot when those clients are given an MCP server. A copy of the block is in mcp/claude_desktop_config.example.json.

ChatGPT

In a custom GPT, add an action from mcp/openapi.yaml. The action posts the design to https://www.reg4rwe.com/api/appraise. In ChatGPT developer mode, you can instead attach the MCP server above. Ask it to appraise the study. The reply should list precedents and the feedback the regulators recorded.

The same POST works for any client that can call an HTTP endpoint:

curl -s https://www.reg4rwe.com/api/appraise \
  -H "Content-Type: application/json" \
  -d '{"design":"A single-arm trial in a rare disease, with an external control arm built from a disease registry."}'

Usage log

Every appraisal is logged: when it happened, whether it came from the MCP server, the hosted API, or a local server, the design text, and the precedents that were returned. The log is not part of the public repository.

The dashboard at www.reg4rwe.com/logs reads the private Vercel usage store. It asks for the USAGE_READ_TOKEN kept in the project environment. Locally, python3 serve.py and http://127.0.0.1:8765/logs read that same store.

The lines on this machine are also in data/usage.jsonl. The hosted copies sit in the private store reg4rwe-usage. A text report is:

vercel env pull .env.local --yes --project reg4rwe
python3 pipeline/usage_report.py

What the layers are

  1. How the evidence was used. Pivotal or substantial evidence, supportive or contextual evidence, safety and labelling, other uses, an unclear role, or no RWE content.
  2. What question it was asked to answer. Post-approval safety and labelling, benefit–risk including external comparators, disease context, support to a clinical trial, other, or none.
  3. The setting. Pregnancy, paediatric use, vaccines, rare disease, REMS, class warnings, oncology. A document can sit in more than one.
  4. What the regulator accepted or criticised. Tags read from the assessment, the methodological issues, and the extracted passage.

Rebuild the corpus

data/corpus.json is the published corpus. pipeline/build_corpus.py rebuilds it from the curated extracts in data/curated/ and the screen in data/search_update/. pipeline/locate_documents.py attaches the public review or assessment-report link and the page. pipeline/update_search.py repeats the public screen and needs pypdf plus network access. Downloaded PDFs stay in data/cache/ and are not part of this repository.

About

Regulatory feedback for real-world evidence, from public FDA reviews and EMA assessment reports.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages