Data Versioning Intelligence

The silent data change, caught.

Every structural check is green — schema, freshness, row count, nulls — and the business number is still wrong. DVI detects the semantic change, attributes it to the deploy, and maps the blast radius.

revenue · dailylast 30 runs
drift begins
schema OKfreshness OKrow count OKnulls OK
Every structural check passes. The value is still wrong.

The pipeline

How it works

01

Profile

Snapshot each column before and after a change — distributions, not just schemas.

02

Detect

Signature detectors flag semantic shifts a structural check can never see.

03

Rank

Corroborate symptoms against the lineage and rank the most likely root cause.

04

Blast radius

Project the change onto downstream assets and the external consumers at risk.

What DVI sees

The signatures

01

Value substitution

A category is silently renamed — "UK" → "United Kingdom".

02

Distribution shift

A numeric column drifts beyond its historical spread.

03

Cardinality change

The set of distinct values quietly grows or collapses.

04

Format drift

Casing or formatting mutates under a refactor.

Evidence

Measured, not marketed

100%
Injected recall
0
Real-vs-real false positives
0.047
Out-of-fold ECE

Validated on 53,940 rows of real data. Detection is deterministic — an LLM may narrate evidence, but never decides whether something changed.

Add DVI to your pipeline.

One command, deterministic detection, a pull-request comment when a number silently moves.

$ pip install dvi
github.com/anuran-de/dvi