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.
Define the evaluation position first
Define whether an even window evaluates at the original sample position before treating value differences as filter error. A central evaluation can be defensible while conflicting with the consumer’s sample-grid requirement. Changing output coordinates therefore needs an explicit contract. Prepare the Truthound reference externally from the polynomial value at the approved evaluation position. [1, 2, 3]
Follow-up discussion of conflicting requirements
SciPy #24512 questions polynomial preservation under an even Savitzky–Golay window. Follow-ups describe a conflict between central evaluation and an asymmetric sample-centered window, rather than declaring every even window defective. A proposed manual pos script explicitly lacks checks for every edge case. [1]
Admit coordinates alongside values
Record window length, polynomial order, and the approved evaluation position. Compare known polynomial values at the required positions with the actual output, withholding results labeled with the wrong coordinates. Operators provide the observed values and independently defined grid reference under the same sample ID. [1, 3]
The difference between 4 and 6.25 in x²
For illustrative f(x)=x² and approved x=2, the required value is 4. An observed 4 passes. The value 6.25 at x=2.5 fails if attached to the original x=2 coordinate. If a different evaluation position is approved, derive the reference at that position instead. [1, 3]
Comparison outside coefficient calculation
ColumnPairValidator in Truthound Python SDK 3.3.2 compares actual filtered values with value_at_approved_grid calculated by the operator. Polynomial references, boundaries, and window positions are external design choices. Core does not calculate filter coefficients or declare every even window wrong. [1, 2, 3]
Choose an odd window or new coordinates
Where sample-centered values are required, use a verified odd window or explicitly retain the actual coordinates of an even-window output. Existing regressions for the same grid requirement can reduce the need for another audit. Recheck boundary handling before treating a positional workaround as general support. [1, 3]
Application conditions
Provide window length, order, evaluation position, and an independent polynomial reference on the approved grid.
Alternatives
Choose a verified odd window or explicitly label the actual coordinates of an even window.
Limitations
This does not declare even windows inherently wrong or verify coefficients and boundary handling.
Evidence
- Event checked
- Subject checked
- Modified
