jev-clean

What it is

An experimental decision-first CSV cleaning tool for data engineers who need auditable, conservative repairs. It profiles issues, builds bounded repair candidates locally, and uses Jev to judge applicability and semantic safety before committing versioned snapshots.

How it uses Jev

Jev Choice selects an action, Noul returns P(applicable), and ordered Score returns P(risk level = 0/1/2) for each candidate. A deterministic policy gate accepts or abstains based on these distributions, expected risk, and snapshot eligibility; accepted candidates are executed and semantically postchecked.

Primitives:choicenoulscore

Technique worth stealing

Separate deterministic candidate generation from model judgment, then gate repairs with fixed policy thresholds and record full distributions.

Try it

uv run idac clean --input examples/data/dirty.csv --config configs/demo.yaml --output runs/my-first-run

View on GitHub