bundle/v0/ESTIMAND_DIFFERENCES.md

# Estimand differences and unavailable inputs

1. **Undisclosed seeds.** The paper does not publish dataset, bootstrap, or OOS seeds. This reproduction uses SHA-256-derived seeds in a fixed namespace and reports distribution-level uncertainty rather than claiming bitwise author equivalence.
2. **Rounded published parameters.** Tables 2 and 3 print rounded Gaussian moments and regression parameters. Full-precision draws are unavailable. Rounded covariance matrices are symmetrized and eigenvalues below `1e-6` are clipped; the Frobenius adjustment is recorded for every task.
3. **Sample-size notation.** The paper writes `J in {10, ..., 100}` and plots ten-sample increments. The frozen plan interprets this as `[10,20,...,100]`; regression figures use `[10,20,30,40,50]`.
4. **Optimizer mismatch in released code.** Algorithm 2/Table 1 specify constant-step gradient descent with raw-gradient clipping at `[-1000, 1000]`, whereas the only public commit unconditionally applies Adam before clipping at `[-10000, 10000]`. The primary route is `paper_algorithm2_gd`; `released_source_adam` is a mandatory labeled implementation sensitivity. Neither route is silently substituted for the other.
5. **OOS Monte Carlo.** Main absolute-regression evidence uses the paper's 10 million samples. Appendix F.3 uses 10 million for its exemplar but states 1 million for the squared-loss multi-experiment trajectory. Those two squared-regression estimands are preserved separately. OOS RNG streams are frozen and separate from training data.
6. **Theorem scope.** Empirical stationarity is supporting evidence only. The completed theorem audit is `HOLD / inconclusive`: the cited diminishing-step result concerns critical accumulation points and objective-value convergence under unverified assumptions, not unconditional full-iterate convergence. Worsening traces alone are not a logical counterexample.
7. **Gaussian CVaR coefficient fork.** The paper's Gaussian risk definition uses `phi(Phi^-1(1-gamma))/gamma`, while the Appendix/official SOCP uses `sqrt((1-gamma)/gamma)`, corresponding to a broader moment class. The paper-scale primary estimand is `gaussian_exact`; `moment_class` is a mandatory sensitivity and is never silently substituted.
8. **Bootstrap quantile.** The paper does not define interpolation. Primary runs use NumPy's linear quantile to match the released implementation; receipts also report the `higher` quantile sensitivity.
9. **Dirichlet concentration.** Appendices F.2.1-F.2.2 say weights are Dirichlet but omit concentration parameters. The scaffold uses symmetric concentration `1.0` and marks these supplementary routes non-exact until clarified.
10. **Stopping-rule fork.** Paper Algorithm 2 stops on signed relative improvement of the total penalized objective `phi`; the released source stops on the non-penalized lower objective and divides by `abs(previous)`. Primary execution records the literal paper diagnostic, an absolute-denominator safety diagnostic, and the released-source diagnostic separately. Worsening-triggered stops are explicitly labeled and are never described as convergence.
11. **Empirical-portfolio algebraic reduction.** Under full support, uniform empirical weights, and linear portfolio loss, sample epigraph variables in Appendix A.2.1 are redundant. The independent implementation removes them while preserving the robust objective and decision.
12. **Squared-regression value representation.** Appendix A.3.2 solves a nonnegative root expression. The scientific squared-loss value is computed by squaring only after the conic solve, and the metric derivative receives the corresponding `2*root` factor.
13. **Bootstrap counts.** The paper exemplars use 20 bootstraps for main Gaussian portfolio and absolute regression, and 10 for squared regression. It does not fully specify every multi-experiment and appendix suite. Counts extrapolated from exemplars are disclosed per suite. The released demos use 5 or 10 and are not the scripts that produced the paper figures.
14. **Iteration display.** The released examples reconstruct stored iterations as `100, 200, ...` even though the optimizer stores iteration 0. Reference validation uses the optimizer's recorded iteration indices, not the reconstructed plotting axis.