Styxis
Company
Knowledge

True in a Dask object column: identical text is not an identical Boolean state

Use original value types and missingness semantics as the handoff contract rather than display text.

The original type behind the text

Use original value types and missingness semantics as the handoff contract rather than display text. Extract Python value kinds before preparing Truthound inputs. A visible True can hide a Boolean-to-string change that alters downstream state handling. [1, 2, 3]

Classification without inspecting values

Dask #12176 reports pandas Boolean and NaN values in an object column becoming strings and NA through Dask. A follow-up points to automatic string conversion classifying a bare object dtype rather than inspecting values. Keep that candidate explanation scoped to the report instead of claiming every object column changes. [1]

Preserve types with masks

Record raw value types, missing masks and actual computed types. Hold consumption if a Boolean contract becomes string before an explicit conversion policy is chosen. Define approved Boolean, string and missing mappings, retaining value-level kinds rather than replacing them with one column dtype. [1, 3]

Two semantic types for True

The illustrative state column permits actual Boolean True and missing values. Its first row requires semantic type boolean; observed boolean meets that condition. Text True has type string and fails. Correct type still requires checking values and the locations and counts of missing records. [1, 3]

A bounded comparison of kind labels

Truthound Core Python SDK 3.3.2 can compare prepared approved_semantic_type and actual_semantic_type using ColumnPairValidator with relationship="==". The operator extracts original value kinds into non-missing labels. The SDK does not change automatic string-conversion settings, and type agreement does not establish equal values or missing counts. [1, 2, 3]

Handoff with an explicit type

Choose explicit nullable Boolean or a verified conversion setting and reconcile computed values and masks again. A field intended as textual state needs another contract. Retain existing raw-type evidence where the same consumer path is already reconciled. [1, 3]

Application conditions

Extract original types without loss and define approved Boolean, string and missing mappings.

Alternatives

Choose explicit nullable Boolean or verified conversion settings and reconcile computed values and masks.

Limitations

  • Core compares semantic-type labels, does not change automatic conversion and cannot prove value or missing-count equality from types.

Evidence

Event checked
Subject checked
Modified