Skip to content

README.en.md

Latest commit

 

History

History
103 lines (75 loc) · 2.89 KB

README.en.md

File metadata and controls

103 lines (75 loc) · 2.89 KB
Raw

ResearchStat AI

AI-native statistical analysis and reproducibility platform for scientific research.

Simplified Chinese is the default project language. The canonical Chinese introduction lives in the root README.md.

English | 简体中文 | 日本語

V1 implements the complete loop:

AI planning
  -> Protocol binding
  -> Human review
  -> Python/R engines
  -> Cross-engine validation
  -> Audit trail
  -> Publication figures

Core Features

  • Protocol Registry: YAML-defined methods, assumptions, posthoc, alpha, missing-data policy, and effect sizes.
  • AI Planner: natural language to a structured plan, restricted to registered protocols.
  • Human Review: accept or override a recommendation and record the reason.
  • Multi Engine: the same request runs on Python and R with cross-validation.
  • Validation: NIST StRD, R official datasets, and boundary data with 1e-8 continuous tolerance and 1e-6 p-value tolerance.
  • Audit Trail: analysis_record.json for every analysis.
  • Figure Engine: editable SVG, PDF, and TIFF at 300 DPI, with scatter, boxplot, violin, and survival plots.
  • Privacy: field masking, identifier hashing, and temporary-workspace cleanup.
  • MCP: stdio, SSE, and streamable-http transports.

Installation

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e ".[figure,test,mcp]"

Quick Start

Do not want to read code? Start with the tutorial (Simplified Chinese), which includes prompts you can copy directly into your agent.

Example data: examples/data/tutorial_data.csv

import pandas as pd
from researchstat.workflow import run_analysis_workflow

data = pd.read_csv("examples/data/tutorial_data.csv")

output = run_analysis_workflow(
    user_input="compare three drugs on mouse tumor size",
    data=data,
    outcome="value",
    group="group",
    audit_dir="audit",
)
print(output["result"].model_dump())

MCP

.\.venv\Scripts\python.exe -m researchstat.mcp.cli --transport stdio
.\.venv\Scripts\python.exe -m researchstat.mcp.cli --transport streamable-http --port 8000

Exposed tools:

  • list_protocols
  • plan_analysis
  • execute_analysis
  • render_figure

Tests

.\.venv\Scripts\python.exe -m pytest -q
.\.venv\Scripts\python.exe benchmarks\run_performance.py

Documentation

  • Project plan: PROJECT_PLAN.md
  • Tutorial: docs/TUTORIAL.md
  • Research lessons: docs/RESEARCH_LESSONS.md
  • Open source ecosystem: docs/ECOSYSTEM.md
  • V1 acceptance: docs/V1_ACCEPTANCE.md
  • Paper draft: docs/PAPER.md
  • Beginner tutorial: docs/TUTORIAL.md

License

MIT. See LICENSE; third-party licenses are listed in NOTICE.

Architecture

See the editable draw.io architecture source and SVG preview.