If the first NaN slice determines NumPy nanquantile dtype: test a reordered handoff
Check final dtype and per-slice quantile references so ordering and missingness do not change approved precision.
When ordering becomes a precision condition
Check final dtype and per-slice quantile references so ordering and missingness do not change approved precision. Supply Truthound with the computed dtype rather than graph or input metadata. One missing slice changing valid slices’ precision would make source ordering an unintended analysis parameter. [1, 2, 3]
Allocation from the first slice
NumPy #32832 reports float32 nanquantile returning float32 when the first slice is all NaN, rounding later slices differently by order. It describes an array-q path and first-slice allocation. Keep the final-dtype and rounding scope rather than claiming every quantile is wrong or a repair has shipped. [1]
Join the reordered control by ID
Preserve slice IDs, NaN masks, q and output dtype; move the all-NaN slice between the first and later positions. Withhold results with an unapproved dtype from precision-sensitive consumers. Join results by slice ID and define the all-NaN result policy separately. [1, 3]
The dtype contract for q=[0.3]
The illustrative policy supplies q=[0.3] quantiles of float32 measurements as float64. An empty first slice does not change that required dtype. float64 meets the handoff type condition; float32 fails. Dtype agreement still requires checking individual quantiles and missing-slice meaning. [1, 3]
Compare actual result signatures
Truthound Core Python SDK 3.3.2 can compare prepared approved_quantile_dtype and computed_quantile_dtype using ColumnPairValidator with relationship="==". The operator extracts the actual result signature and preserves slice correspondence. The SDK does not repair first-slice allocation, and dtype agreement does not guarantee quantile values. [1, 2, 3]
Values a later cast cannot recover
Test explicit separation of valid and all-NaN slices or construction of an approved output dtype. Casting an already rounded result to float64 does not recover lost precision. Hand off a path checked against both reordered slices and numerical references. [1, 3]
Application conditions
Read dtype from the actual computed result, joining reordered output by slice ID and defining all-NaN meaning separately.
Alternatives
Test a verified valid/all-NaN separation or an explicitly approved output construction.
Limitations
Core compares extracted dtype signatures, does not repair allocation and cannot prove quantile values from dtype alone.
Evidence
- Event checked
- Subject checked
- Modified
