ivyxivyx
← All skills
IVYXSTUDIO · SKILL
D

Data contract authoring

data-contract-authoring · v1.0.0

ivyx

Use when someone asks for a data contract, a pandera schema or a validation gate (veri kontratı, şema kontrolü, pandera, doğrulama) for a training file, or wants the pipeline to refuse bad data before training (eğitimden önce bozuk veriyi reddetsin); drafts the contract from a profile and adds it to the pipeline as a gate stage. Not for a first look at a file.

Who is asked

Risk medium: you are asked when it loads, and that yes covers reading its files in the same chat.

Tools it may call

While this skill is active, a call to any other tool is refused. These still meet every policy rule; the list narrows, it approves nothing.

duckdb.inspectprofiler.profilemlops.pandera.contractFromProfilemlops.pipeline.describemlops.stage.add

What it tells the agent

The published SKILL.md, as written.

---
name: data-contract-authoring
description: Use when someone asks for a data contract, a pandera schema or a validation gate (veri kontratı, şema kontrolü, pandera, doğrulama) for a training file, or wants the pipeline to refuse bad data before training (eğitimden önce bozuk veriyi reddetsin); drafts the contract from a profile and adds it to the pipeline as a gate stage. Not for a first look at a file.
allowed-tools: duckdb.inspect profiler.profile mlops.pandera.contractFromProfile mlops.pipeline.describe mlops.stage.add
metadata:
  risk: medium
---

A data contract for a training file

A contract is written from what the data measurably is, then added to the pipeline as the first gate. You do not write the contract file or edit .punica/mlops.yaml by hand: the two capabilities below write them, and the person says yes to each write.

1. Measure the data

  1. duckdb.inspect with the file's path as file: columns, row count, the first rows.
  2. profiler.profile with { "path": "<file>" }: per column, the type, the null share, the distinct count, and for numbers the min and max. Keep the result; the next step takes it as its input.

2. Draft the contract from the profile

mlops.pandera.contractFromProfile with the profile you just got (or the file's path, which profiles again), contract as the path to write (contracts/<file-stem>_contract.py unless the person named one), and:

  • ranges: true when the person wants bad values refused, false when they want only the columns and types checked. Say which you chose and why.
  • Leave strict off unless they said a new column should fail the check.

Read the answer's columns and notes: a column the profile saw with nulls, a label column with few distinct values, a range that looks like a measurement limit. Report them as sentences, with the measured numbers.

Then read the draft against the data: the draft accepts every value the file holds today, so a value that is already wrong in the file passes it. For the label or class column, list its distinct values from the profile; if the draft checks it with a range (in_range(0, 2)) rather than by its allowed values, say so in your report and name the line the person should change (pa.Check.isin([0, 1]) with the values they consider valid). You do not edit the file yourself.

3. Add it to the pipeline

  1. mlops.pipeline.describe: the stages in order and the engine.
  2. mlops.stage.add with capability set to the pandera validate gate's id as the describe output or the catalogue names it, id: "pandera", title, onFailure: "stop", and input holding contract (the path written in step 2) and data (the file). It is placed before the training step; read order from the answer and say it.

If a contract stage already exists, do not add a second one; report the existing one and the contract it points at.

4. Report

  • The contract path, the columns it checks and with what (type, nulls, range, values), each with the number it came from.
  • The stage: its id, where it sits in the order, what a failing run will say.
  • What to look at before trusting it: a range taken from a small sample, a column the profile found empty.
  • Offer to run the pipeline once the person is ready; do not run it yourself.