Time weights: default tuning,
different optimization
The same SDID samples and time-weight penalty, solved with different numerical settings. No penalty retuning or cross-validation. Original results are preserved.
Inspect a cohort and outcome
The two charts share a y-axis scale within each comparison. Gray bars are at or below the support cutoff; the underlying vectors are never thresholded or renormalized by this control. Hover bars for source-period labels and weights.
Inspect the selected weight vectors
One preselected example per paper
The same representative examples as in the default report; click a paper to inspect it. Support uses weight > 10⁻⁶. The available papers, outcomes and cohorts are not an independent or representative sample of studies.
How to read these comparisons
- Defaults: author sparsification and 10,000 final iterations; improvement tolerance 10⁻⁵ × noise scale.
- Author tighter: 100,000 final iterations and tolerance 10⁻⁸ × noise scale, with and without sparsification.
- Optimized reference: the identical full-simplex convex objective, not a different SDID penalty. The independent objective-gap target is 10⁻⁸ after normalizing by noise variance. All 11 nonzero-noise opera blocks receive an additional 80-digit active-set solution; other blocks retain the double-precision QP. All original solver variants remain available.
- Near-flat directions: the normalized ridge coefficient is only 10⁻¹². Small objective gaps alone do not prove the weights themselves are accurate or stable. The high-precision check resolved three initializer-sensitive opera cases and refined two more whose double-precision starts had agreed. Outside opera, passing objective and initializer checks is not a universal certificate of weight accuracy.
- Numerical zeros: QP solutions can have tiny roundoff entries. Raw weights and the small feasibility repair are exported; strictly positive support is not directly comparable across solvers.
- Scope: time weights only. No unit-weight refits, new ATTs, bootstrap or corpus expansion. Two zero-noise blocks remain without substitute weights. Government audits are randomized; existing pending-design labels are unchanged.
Findings and comparison profiles · Base numerical report · High-precision refinement · Profile PDF · All five variants’ vectors · Diagnostics · Paired comparisons · Optimized reference vectors · Initializer checks · Support thresholds · Quarto source