Pillar 03 · Automation
A1 Fitness-for-AI Check Library
Rules that run where data moves keep pace with AI.
Form Workbook and YAML
What it's for
An AI agent can use bad data thousands of times before a quarterly review notices. This library gives you checks as code: completeness, uniqueness, validity, schema, freshness and volume, plus the fitness-for-AI checks the framework adds: decision freshness, authority, leakage, representativeness and bias. Every check is written in SodaCL in the accompanying file, gfa_fitness_for_ai_checks.yml.
The checks were tested with Soda Core 3.5 against sample data. A failed Critical check makes the scan exit with code 2, which your orchestrator can use to stop the load.
What you get
- Catalogue The 32 checks, each with what it protects against, a severity, threshold guidance and its SodaCL, and your choices: adopt or not, your dataset, threshold, owner and severity.
- Coverage Checks in the library, the critical ones and those you plan to adopt, by dimension, with the gap.
- Dimensions The question each dimension answers.
Worked out for you Checks in library, You plan to adopt, Gap
How it connects
Choose the feeds from S1 and S2; each check's severity follows A3, and A2's contracts point their quality thresholds at the checks you adopt. C4's flagged answers often end in a new check.
A flavour of the worked example
| Check | Rent doesn't jump between periods |
|---|---|
| Dimension | Validity |
| Your dataset | fact_rent_roll |
| Your threshold | Less than 50% from last period |
| Your severity | Critical |
A few fields from one record of the toolkit's worked example, with long text shortened.
Next step
Guided build coming next.
Request access
Ask for access to the guided builds, including this kit's when it is ready.