dataproof

Find out what your dataframe code actually did.

import pandas as pd, dataproof

with dataproof.observe() as report:
    orders = pd.read_csv("orders.csv")
    clean  = orders.dropna()
    joined = clean.merge(tiers, on="customer_id")
    gold   = (joined.tier == "gold").sum()

print(f"gold customers: {gold}")
print(report)

gold customers: 0
dataproof: 1 finding(s)
  - dropna() removed ALL 122 rows — a row was dropped if any value in it was
    null. Pass how='all', a subset=[...] of columns that matter, or thresh=N
    to keep partly-populated rows.  [analysis.py:5]

The answer was zero. The real answer was 33. One column in that file was entirely empty, so every row contained a null, so dropna() removed all of them — and pandas said nothing, because nothing went wrong as far as pandas is concerned.

That is the problem this solves. Not crashes: crashes are easy. Plausible answers that are quietly wrong.

What it watches for

What it does not do

It does not change your results. Every wrapped call returns exactly what pandas returned. The code inside the block is ordinary pandas.

It does not ask you to write differently. No new API to learn, no wrapper types, no rewriting a pipeline to use someone else's functions. That is deliberate — a verification tool you have to remember to call is one you will not call.

It does not cry wolf. Every threshold here exists because a noisier version was tried and rejected. A group-by is not a 96% row loss. A few points beyond the IQR fence are not outliers worth reporting unless they actually move the mean. A finding that fires on correct code teaches you to skip all of them, which is worse than having none.

Install

pip install dataproof

Pandas is the only dependency.

Also useful

Check any frame on demand, without a block:

dataproof.Report().check(df)

Use it as an assertion in a pipeline or a notebook:

with dataproof.observe() as report:
    ...
assert report.ok, str(report)

Why this exists

It came out of building a data library for AI agents, and measuring what agents actually do. Given six purpose-built tools and a free choice, an agent used none of them and wrote pandas — correctly. The tools were not the problem. The problem was that neither the agent nor the person reading its answer could tell when an operation had quietly destroyed the data.

Agents now write a great deal of dataframe code, at a speed no one reviews line by line. They are good at it. But dropna() behaves the same for everyone, and the failure is silent for everyone.

MIT licensed.