Truthound · Numerical and statistical validity · 2/3
Separate finite values, units and precision from statistical validity.
58 articles
Check original MILP constraints before interpreting optimal status
Separate feasibility from optimality in that order. Do not use equality to one solution vector as the only reference.
Admit the value remaining after signed logsumexp cancellation
Use an independent reference for the remainder and sign after leading terms cancel. Do not replace NaN with an arbitrary zero.
A SciPy Laplacian diagonal of -56: a correct degree written into a narrow buffer
Even a correct degree of 200 can be lost when assigned to a narrow work matrix. Compare the original-edge reference with the actual diagonal.
Before clipping a tiny negative interpolated variance at an input knot
Check knot reproduction separately from variance nonnegativity. Even tiny values need an error policy appropriate to their use.
Do not dismiss an integer matrix’s 0.5 power as a dtype preference
Use a squared residual to check preservation of the fractional exponent. A plausible dtype can conceal a different operation.
Which grid does an even Savitzky–Golay window describe?
The central evaluation point and the original sample grid can be different contracts. Do not read a half-grid shift as error magnitude alone.
Explain a tiny slope in an exactly constant series
Retain direct evidence of constancy and the task’s error budget. Separate a tiny slope from the interpretation of its p-value.
When reversing Canberra distance inputs changes the sign
A finite distance alone cannot admit a neighbor relation. Compare both directions with a reference derived from the original vectors and preserve the integer absolute-value boundary.
Before inheriting UTC sortedness for local calendar dates
A calendar reversal after timezone conversion differs from UTC order. Derive the minimum from actual values rather than trusting the flag.
Do not admit a large Float64 remainder from engine agreement alone
Build an independent reference from the integer identity. Numerical-path differences can arise even without NaN inputs.
Check observation presence when a singleton GPU datetime mean is NULL
Use the identity that one valid instant has that same mean. Distinguish operation support from CPU-fallback requirements.
Retain intermediate meaning when a Boolean cast chain produces -24
Check the nonzero-is-true policy beside the final 1/0 codes. An Int8 dtype alone cannot admit the status code.
A NaT position from NumPy nanargmin: inspect the original mask before handing off the minimum
Check the selected position together with its original validity so a NaT index is not handed off as the minimum observation.
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.
NumPy masked comparison at 0.7: agree on dtype and a strict boundary first
Approve comparison precision for observations and thresholds before checking masked and ordinary paths.
NumPy gradient on integer nanoseconds: preserve the origin and intervals before float conversion
Preserve integer differences before converting large absolute times to float, then approve gradients against the original intervals.
NumPy einsum optimize: was casting=safe enforced on the final out write?
Verify rejection of a forbidden final cast so optimization cannot change permission for data loss.
Large offsets in NumPy cov: separate positive covariance from an approved error budget
Even positive covariance can lose required accuracy at a large offset.
NumPy corrcoef on constant channels: do not rank a finite 1 as defined correlation
Pearson correlation is undefined for a constant channel.
Rescaling complex sqrt on Windows: a correct real part does not validate the imaginary part
Check both components when restoring scale for large complex inputs.
