Navigation
DashboardProbesRealness Loop › Iter 26 · #13 slice 1

iteration 26 · 2026-07-24 · conditional axis · bounded slice 1 · laptop-drives-bigblack

➗ #13 Double/Debiased ML + CPI CHECKPOINT

The third conditional instrument: does the feature carry a forward-return effect that survives orthogonalisation against the shipped set Z? Cross-fitted DML partialling-out + a HAC standard error + the CPI predictive-impact test. All four slice-1 gates pass, robust across 2 seeds — a clean slice, because DML's orthogonalisation handles the very dependence that broke the knockoff FDR in #12.

0.94
DML CI coverage (target .95)
0.967
admit power (HAC-t 6, CPI p .003)
1.22 / 0.24
retention: indep vs Z-explained
2 seeds
both 4/4
Preflight (resource-only): load1 1.77 · 37 GiB available · si/so ~0 · ClickHouse active, readonly=2. 5c/5G/no-swap capped; single-thread BLAS. Reused the #12 wide slice for the conditioning set Z.

The four slice-1 gates

Gate (§7 row 13, slice 1)Result (seeds 20260723 / 11)Target
Gaussian coverage & null t~N(0,1)coverage 0.94 · null-t mean 0.18 std 0.91 · FP 0.05 / 0.03coverage ∈[.93,.97] & FP≤.05PASS
Admit (R≥.5, HAC t≥3, CPI p<.05)power 0.967, R 1.05, HAC-t 6.0, CPI-p 0.003power ≥ .8PASS
Null FP (block-perm target)0.067, null-t std 0.93≤ .06PASS
Retention calibrationindep 1.22 (admit) · Z-explained 0.24 (reject)R_hi>.5 & R_lo<.5PASS

The two axis controls

Coverage / calibration. On a jointly-Gaussian construction with a known partial effect θ* (Y=Z·w_Y+e_Y, f=Z·w_f+θ·e_Y+e_f → exact partial slope θ/(θ²+1)), the cross-fitted DML confidence interval covers θ* in 94% of reps and under θ*=0 false-positives at 5% — the estimator is calibrated. (The null-t mean 0.18 is a mild bias from cross-fitting on autocorrelated Z; coverage + FP are the operative calibration and both pass.)

Retention separates conditional-real from Z-redundant. A feature whose signal is mostly Z-independent retains R≈1.2 (admit); a feature whose signal is mostly Z-explained — genuinely sharing Z's contribution to a Z-dependent Y — collapses to R≈0.24 (reject). The R=.5 threshold cleanly separates a feature that adds forward information from one redundant with the shipped set.

The conditional axis, three ways. #11 CMI (information-theoretic conditioning, GROUNDED) · #12 knockoffs (FDR-controlled selection, operator-ruling — the real long-memory panel broke exact FDR) · #13 DML/CPI (orthogonalisation) — here the HAC on the residual regression absorbs the autocorrelation that broke #12, so the slice is clean. Slice 2 completes it: ABSTAIN@corr(f,Z)≥.95 (collinearity guard), HAC-necessity (iid SE fails @ Hurst≥.7 — the #4/#5 long-memory point), and the incomplete-Z Harden (omitting a true premium from Z keeps R spuriously high → Z-PREREG + sensitivity-as-Z-grows + unobserved-confounder null).