Data invariants¶

Structural guarantees on the validation data itself. These tests never call the model — they validate the data contract upstream of any modeling.

DataInvariantTest¶

Assert expected columns are present and that no column has too many missing values.

from modeltest.scenarios import DataInvariantTest

DataInvariantTest(expected_columns=["age", "income"], max_null_ratio=0.02)

Param	Type	Default	Description
`expected_columns`	list[str] \| None	`None`	Columns that must exist in `X_val`; `None` skips the column check.
`max_null_ratio`	float	`0.02`	Maximum tolerated fraction of nulls per column. Set negative to disable the null check.

- type: data_invariant
  params: {expected_columns: [age, income], max_null_ratio: 0.02}

Failure details: Missing columns: ['income'] or Columns exceed null ratio 0.02: {'phone': 0.31} (it names every offending column and its ratio).

NoNullTest¶

Assert no null values exist in the validation data (or in the selected columns).

from modeltest.scenarios import NoNullTest

NoNullTest()                       # whole frame
NoNullTest(columns=["age", "income"])  # subset

Param	Type	Default	Description
`columns`	list[str] \| None	`None`	Columns to check; `None` checks the entire `X_val`.

- type: no_null

Invariants vs drift

Invariants are hard schema/quality rules (schema changed, feed broke); drift tests are statistical comparisons (distribution moved). Use invariants to catch pipeline breakage fast, drift to catch slow decay — see Drift.