Truthound · Numerical and statistical validity · 3/3
Separate finite values, units and precision from statistical validity.
58 articles
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.
Declare timestamp precision before a DuckDB Parquet glob chooses it
Compare actual timestamp values when changing the first file in a glob. An output dtype alone can make already-rounded timestamps appear acceptable.
Check NaN constants before treating negated DuckDB comparisons as algebra
Reconcile row-selection flags for the original negated comparison and its optimized path when NaN is present. Quarantining every non-finite value does not establish that the original predicate was preserved.
Cross platform and list-length boundaries for DuckDB BIGNUM lambdas
A scalar control is insufficient to approve lambda comparisons involving a large negative number. Check result Booleans across list lengths and element positions on the operating system that will run the query.
Use an independent rounding reference for Doris DECIMAL casts
Build a decimal-cast reference with an independent parser reading directly into the target type. An intermediate DOUBLE or a shortened display can conceal a second rounding step on the FLOAT path.
Align the Doris ANALYZE count and load-counter baseline
Align the COUNT snapshot and load-counter generation before adding a load delta to ANALYZE statistics. A race in the estimate does not establish duplicated stored rows.
Preserve the kind of nonfinite value after ClickHouse rounding
A finite check cannot distinguish NaN from infinity after rounding. Retain input and output class tags to preserve the difference between an indeterminate calculation and an infinite magnitude.
Choosing a NaN policy before moving Polars comparisons to a GPU
When NaN > 0 selects different rows across engines, matching Boolean schemas are not enough.
Checking Polars Decimal shift fill values when 1.25 becomes 0.01
Decimal accuracy in existing rows and scale conversion in newly filled rows are separate contracts.
DuckDB integer overflow: do not rely on WHERE to protect multiplication
Success that depends on filter evaluation order is not evidence that arithmetic inputs are safe.
A DuckDB SUM that cannot fit UINT8: validate aggregate output capacity
Small integer inputs can require a wider aggregate output type.
Pydantic Decimal multiple_of: sharing exact counterexamples beyond the precision boundary
Teach and verify exact divisibility through independent arithmetic on the original decimal representation, rather than a rounded quotient.
jq large-integer subtraction: keeping raw numbers in the team handover
A small difference can still be wrong when its inputs are large integers. Preserve the original numbers because precision may be lost before subtraction.
Dask versus NumPy cumulative sums: define an error budget before judging chunk boundaries
Interpret floating-point cumulative-sum differences against an independent reference and an agreed error budget.
Validate Dask complex Cholesky with conjugate-transpose reconstruction residuals
Check complex matrix factors against the original equation rather than copying a transpose check from a real-valued example.
Check quantile probabilities before interpreting Arrow tdigest NaN
Invalid probability requests and approximation error on valid requests require different checks.
Checking Arrow rounding grids separately from tie-breaking
Verify both membership of the rounding grid and the policy used to choose a neighboring multiple.
Checking the int64 minimum before negating Arrow durations
The smallest signed duration has no positive counterpart in the same storage type.
