Loading...
Asserts that the arithmetic mean of a numeric column falls within an expected range. Detects data drift, calculation errors, or upstream changes that shift the central tendency of key metrics.
Validates that feature distributions in production data do not deviate more than the configured number of standard deviations from the training baseline. Data drift indicates that production inputs have shifted from what the model was trained on, potentially degrading AI system performance. Under EU AI Act Article 10, ongoing data governance requires monitoring for dataset relevance.
Asserts that the null rate of a column has not changed more than the specified percentage points from a known baseline. Detects regressions in data completeness that may indicate broken upstream transformations, schema changes, or ETL failures.
Asserts that the percentage of values falling outside 1.5 times the interquartile range (IQR) is below a specified threshold. The IQR method is a robust, non-parametric approach to outlier detection that works well across different distributions.
Asserts that the Pearson correlation coefficient between two numeric columns falls within an expected range. Detects broken relationships between features, data pipeline errors that decouple naturally correlated fields, or upstream schema changes.
Asserts that the value distribution of a categorical column matches expected proportions within a given tolerance. For example, verifying that 60% of records are type A, 30% type B, and 10% type C. Detects upstream labeling changes, data skew, or ingestion bias.
Validates that the rate of statistical outliers (values exceeding 3 standard deviations from the mean) in a feature column stays below the configured threshold. Excessive outliers in training data can skew model learning and produce unreliable AI systems. Under EU AI Act Article 10, training data must be free of errors to the best extent possible.
Asserts that the standard deviation of a numeric column has not changed more than the specified percentage from a known baseline. Detects changes in data variability that may indicate corrupted data, changed source systems, or process failures.
Validates that the class distribution in training data does not have an imbalance ratio exceeding the configured threshold between any two classes. Under EU AI Act Article 10, training datasets must be representative and free from bias. Severe class imbalance can lead to biased model predictions and underperformance on minority classes.
Asserts that the number of distinct values in a column falls within an expected range. Detects issues such as collapsed categories (too few distinct values), data explosion (too many), or enum drift from upstream changes.
Asserts that the median of a numeric column falls within an expected range. The median is more robust to outliers than the mean, making this check ideal for skewed distributions or data with extreme values.
Validates that protected attributes (such as gender, age_group, ethnicity) are represented with a minimum coverage percentage per group. Under EU AI Act Article 10, training data must be examined for possible biases that are likely to affect the health, safety, or fundamental rights of persons. Insufficient representation of protected groups can lead to discriminatory AI outcomes.
Auto-merged records must carry a match-confidence score at or above the auto-merge threshold; lower scores require manual review.