py-tetrad: Skill for Claude Code

Tetrad Real-Data Analysis

A workflow for applying Tetrad's causal discovery algorithms to real data without producing confidently wrong answers. The full guide, with a complete worked example (Auto MPG) and the finding-code reference, is at TETRAD_ANALYSIS_GUIDE.md in the py-tetrad repository root — read it when you need the details behind any step below.

The principle (non-negotiable)

Agents assist; algorithms discover; scientists judge (Zheng, Verma, Gill, Dai, Spirtes & Zhang 2026, arXiv:2606.23608). You may audit data, explain assumptions, propose methods and settings, run tools, and interpret outputs. You may NOT supply edges, orientations, priors, or causal conclusions from your own knowledge, and nothing you believe about the domain may enter the discovery core except as an assumption the user has explicitly adopted.

Rules of engagement. Background knowledge (tiers, forbidden/required edges), variable exclusions or mergers, type reassignments, transforms, missing-data policies, and parameter choices all change what the algorithm treats as input, so they are user decisions: propose them with reasons, apply them only after the user approves, and record every adopted decision in the final report. Never silently drop or recode a variable, convert a temporal hint into a knowledge tier, or tune a threshold until the output looks better. Present CPDAGs and PAGs as equivalence-class objects conditioned on stated assumptions, never as confirmed causal facts.

The workflow: Audit → Decide → Search → Diagnose → Report

Never run one algorithm with default settings and report the graph.

0. Provenance (ask the user; no software can answer these)

What does each variable mean and in what units? How was the sample collected (selection effects)? Is there a defensible partial time order? What plausible common causes are NOT measured — name them? Did the data-generating regime change during collection? Write the answers down; they become the Decisions section of the report.