Do not interpret a nonfinite singleton variance as stable zero
Keep the population-variance result of a non-finite singleton separate from a finite singleton's no-variation state. Retain input classes and result tags so indeterminate arithmetic is not published as a stability metric.
A finite singleton differs from NaN
The population variance of a finite singleton and the result of a NaN or infinity input are different states. Truthound Core can compare result-class tags without erasing that distinction through a blanket finite-value filter. [1, 2, 3]
A one-row nonfinite report
DuckDB issue #28329 reports var_pop and stddev_pop returning 0.0 for a single NaN or infinity in 2.0.0, while two-row nonfinite groups return NaN. This differs from a sample-variance NULL. A follow-up links a fix PR but does not establish deployment. [1]
Join estimator and input state
Join input finite state, valid count, estimator, and result class by group. Preserve finite-singleton controls and nonfinite-singleton controls. Do not publish the latter’s 0 as evidence of stability. [1, 3]
NAN_STATE versus NUMERIC_ZERO
Define input [NaN] under the declared NaN-propagating population policy. Expected class=NAN_STATE; that passes. NUMERIC_ZERO fails. A separate finite [1.0] population control legitimately expects NUMERIC_ZERO, so the same tag is not universally invalid. [1, 3]
Do not filter away the intended state
ColumnPairValidator in Truthound Core Python SDK 3.3.2 can compare prepared non-null state tags. The operator classifies inputs and supplies estimator policy. A FiniteValidator-only filter would remove the intended result state rather than verify it, and the SDK does not repair aggregate finalization. [1, 2, 3]
Keep the input warning
Retain a separate input-state warning and compare a verified alternative computation’s state. Test the installed version if adopting the linked correction instead of treating issue closure as a release claim. [1, 3]
Preserve the reason for exclusion
State preservation matters even when a downstream report accepts only finite numbers. Preserve the reason a group was excluded so unknown arithmetic does not become an apparently precise no-variation result. [1, 3]
Application conditions
Join estimator, valid n, input finite state, and result class.
Alternatives
Keep input warnings and compare states from a verified alternative computation.
Limitations
Removing nonfinite results alone does not preserve NaN-propagation meaning.
Evidence
- Event checked
- Subject checked
- Modified
