confoundr 0.1.0

A causal validity linter for ML pipelines

Confoundr

A causal validity linter for ML pipelines, plus a deployed platform that runs it as a service

Overview

Data quality tools (Great Expectations, Evidently AI, Feast validations) check whether your data is clean (nulls, schema drift, distribution shift). They do not check whether your data is causally valid — whether the assumptions your causal or treatment-effect model depends on (no leakage, no unmeasured confounding, positivity, balanced treatment groups) actually hold.

Confoundr is an open-source library that runs a battery of causal-validity checks against a dataframe, feature store, or pipeline stage, explains failures in plain language, and — as a hosted platform — lets a team plug in a dataset and get those diagnostics without writing any code, on a schedule, with alerts.

Architecture Highlights

Confoundr is built as a two-layer system:

• Core Library (confoundr pip package): Standalone, dependency-light, usable in anyone's pipeline or CI.
• Deployed Platform: A multi-tenant service wrapping the core library that runs asynchronous checks on uploaded datasets, with observability and an AI explainer layer.

Quick Links

• Architecture
• Check Catalog
• Project Plan