Initial release numerical review
The repaired implementation was inspected again against the methodological contract. This was a source review and a set of independent calculations by the implementation team. It was not an independent human statistical review. The candidate remains available for that review before a stable release.
Consequential findings and disposition
Finding |
Disposition |
Evidence |
|---|---|---|
Power windows could omit an eligible geo entirely |
Reject that replicate and retain its failure |
|
Documentation index still claimed causal proof and a 10–30% optimisation benefit |
Remove unsupported claims and link evidence limits |
|
Package and requirements files specified conflicting dependencies |
One dependency contract in |
Clean installed-package checks |
Tests imported the source layout rather than the public package |
Use |
Full suite outside the checkout |
Broken-reference warnings were suppressed in Sphinx |
Remove suppression and repair unresolved references |
Documentation verification |
Tracked generated HTML retained old claims and embedded old source |
Remove stale build output; regenerate documentation separately |
Tracked-source inspection |
BCa and stratified DiD lacked calibrated release status |
Add explicit experimental result metadata and warnings |
Frozen release simulation |
Mathematical cross-checks
tests/test_release_references.py supplies references that do not call the
production estimator to construct their expected answers.
Public Lift and DiD agree with separately calculated stratified mean contrasts, variance, Welch–Satterthwaite degrees of freedom and t intervals.
Enumeration over all 60 allocations in strata of four and five geos verifies unbiasedness under heterogeneous potential outcomes. Expected Neyman variance equals exact design variance plus the unidentifiable effect-heterogeneity term.
The exact sharp-null p-values match exhaustive enumeration and control rejection probability at the checked thresholds under the permitted assignment design.
BCa bounds agree with an independently assembled within-arm/stratum resampling distribution and delete-one jackknife acceleration. This establishes numerical agreement, not reliable coverage in small samples.
Simulation generator checks verify that treatment leaves baseline outcomes unchanged and that reported truth uses finite-population effects and realised pre-treatment denominators.
The preceding regression suite also checks changed-data analysis, float64 precision, invalid panels, CRT rank deficiency, bootstrap multiplicities, confidence settings, exact and Monte Carlo randomisation, diagnostics, reporting units and all three CLI analysers. See the verification record for checks actually executed on this candidate.
Remaining methodological limits
The Welch–Satterthwaite reference is approximate. The finite-population variance calculation is conservative in expectation under heterogeneous effects; that identity alone does not establish finite-sample interval coverage.
BCa numerical correctness did not prevent undercoverage in the release study. It remains experimental. Stratified DiD also remains experimental because one core coverage screen was inconclusive. Do not change the estimand to a passing method after viewing outcomes.
CRT uses period-adjusted OLS with CRV1 and a t reference. Calendar adjustment, full rank and successful bootstrap draws do not certify small-cluster inference or identify delayed effects. No CRT or staircase calibration was performed in this release phase. Missingness, non-adherence and interference need separate design and analysis decisions.
Primary references
SciPy bootstrap contract describes resampling, BCa intervals and degenerate-data limits.
Pustejovsky and Tipton on small-sample cluster inference explains why conventional clustered Wald tests need scrutiny with few clusters.
Python packaging metadata guidance supports the consolidated package contract. Existing proprietary licence terms are preserved; no public licence is granted by preparing an archive.