iteration 27 · 2026-07-24 · conditional axis COMPLETE · laptop-drives-bigblack
The conditional axis closes. #13 Double/Debiased ML + CPI is GROUNDED — all seven gates, robust across five seeds — the orthogonalisation approach that cleanly survives the long-memory dependence. And per operator ruling, #12 model-X knockoffs is REJECTED: standard knockoffs don't hold on our data; their FDR-selection role is routed to #13.
readonly=2. 5c/5G/no-swap capped; single-thread BLAS. Reused the #12 wide slice for Z.| Gate (§7 row 13) | Result (seed range) | |
|---|---|---|
| Gaussian coverage & null-FP | coverage 0.935–0.965 · null-FP 0.02–0.065 | PASS |
| Admit (R≥.5, HAC-t≥3, CPI p<.05) | power 0.967, HAC-t 6.0, CPI-p 0.003, R 1.05 | PASS |
| Null FP (block-perm target) | 0.067 | PASS |
| Retention calibration | indep R 1.22 (admit) · Z-explained R 0.24 (reject) | PASS |
| ABSTAIN @ collinearity | R²(f\|Z)=.97 → abstain 1.0 · R²=.64 → abstain 0.0 | PASS |
| HAC necessity | iid-SE FP 0.30–0.57 (10–15× nominal) ⇒ HAC necessary; HAC halves it | PASS |
| Incomplete-Z Harden | complete-Z R 0.007 (reject) · omit-confounder R 1.0 (spurious admit), monotone | PASS |
Ruling (Nasim, 2026-07-24): retire #12 — "standard knockoffs don't hold on our data" — and route the conditional-axis FDR-selection role to #13 DML.
Across slices 1–4 the standard Gaussian model-X construction was exhausted: the real panel is near-singular (needs MVR+LW) and, decisively, model-X knockoffs deliver an irreducible realised FDR ≈1.3× nominal (conservative-q calibration fails — no operating point with FDR≤.10 and power; the ~1/q detection floor blocks a smooth tradeoff, and the long-memory + ill-conditioning leave leaking nulls with signal-like W that no threshold separates) with a power ceiling 0.67 @IC=.03. The core FDR guarantee is not met on the real long-memory panel.
Not discarded: #12's interaction-only non-linear-W differentiator (RF 1.0 vs lasso 0.0, slice 2) was genuinely grounded and is documented. Re-test condition: a construction that controls realised FDR≤.10 on the long-memory panel (e.g. an HMM/deep-generative TSKI) — future work, out of scope.