Legacy SHT audit ledger#
This document records the baseline findings that must be resolved before an
affected method can ship in pySHT. The executable baseline is SHT 0.1.9 at
Git commit 4e29cda1257f86dd0237d37329af358b54d04f2b.
The legacy namespace contains 54 statistical tests and two dynamic adapters. The adapters are unnecessary in Python because ordinary callables can be passed directly. Implemented and independently validated slices are recorded separately in the Biswas–Ghosh, classical mean, and classical variance ledgers.
Legacy parity is not evidence of correctness. A method is release-ready only after its primary formula, calibration, numerical implementation, invariants, and null behavior have been independently checked.
Legacy method |
Initial status |
Confirmed concern |
Required gate |
|---|---|---|---|
|
Red |
Dimensionally inconsistent scale-dependent normalization |
Re-derive from Li–Chen and prove common-scale invariance |
|
Red |
Computation ignores the whitened data and supplied null covariance |
Whitening-equivalence oracle |
|
Red |
Equality-facing API implements one directional tail |
Re-derive all supported alternatives |
|
Red |
Reversed p-value tails |
Null calibration and power in both directions |
|
Red |
Legacy asymptotic variance changes under row permutation |
Order-invariant derivation; corrected permutation implementation first |
|
Red |
Stale loop index reuses the wrong group |
Group-order and unequal-size tests |
|
Red |
Raw positive ratio is incorrectly calibrated as standard normal |
Primary-paper re-derivation |
|
Red |
Exact quadrature is unstable and underflows in stronger alternatives |
Log-domain/high-precision oracle |
|
Red |
Documented default call fails |
Default-call smoke tests |
maximum pairwise BF kernels |
Red |
Double-to-float truncation and exponentiation overflow |
Float64, log-domain implementation |
Classical one- and two-sample t tests, scalar chi-square/F variance tests, one-way ANOVA, and Shapiro–Wilk matched trusted baseline implementations in the initial fixed-data probes. These procedures have now passed their pySHT validation gates and are available in their scientific category modules.