iteration 24 ยท 2026-07-24 ยท conditional axis ยท bounded slice 3 ยท laptop-drives-bigblack
๐๏ธ #12 Model-X Knockoffs โ slice 3 CHECKPOINT
Executes the slice-2 correction (wide panel + FDR-controlling selection) and lands on the last obstacle โ the same one seven other instruments hit. The knockoff power contour is shifted right (โฅ.8 at ICโ.04, not .03) and the FDP runs ~1.5ร the nominal q on the real dependent panel: PATTERN #2, the campaign's most-reused fix. Terminal via conservative-q calibration in slice 4.
.71โ1.0
power IC .03โ.07
~1.5ร
FDP vs nominal q=.10
slice 3
conservative-q next
Preflight (resource-only): load1 1.74 ยท 35 GiB available ยท si/so ~0 ยท ClickHouse active, readonly=2. 5c/5G/no-swap capped; single-thread BLAS. Wide 24-feature panel staged read-only.
The power contour (KnockoffFilter, 14 signals, q=.10)
| injected IC | power | FDP |
|---|
| 0.03 | 0.714 | 0.182 |
| 0.05 | 0.964 | 0.149 |
| 0.07 | 1.000 | 0.140 |
| 0.10 | 1.000 | 0.122 |
| no signal (null) | โ | 0.0 selections |
Power โฅ .8 at IC โ .04 (shifted right vs the marginal instruments โ #4 rank-IC is power 1.0 @IC=.03 โ because knockoffs trade power for the stronger conditional-FDR guarantee). FDP is persistently ~1.2โ1.8ร the nominal q at every IC, while the null-only case selects nothing.
Three findings this slice
- The real panel is severely ill-conditioned. The full 24 microstructure features are numerically singular (ฮป_minโ0, cond 8.8e16 โ dead filter). A greedy 20-feature well-conditioned subset (ฮป_min 0.06, cond 68 โ dropping the near-exact duplicates turnoverโกofi, intra_ofi/intensity/kyle) is the usable panel.
- Construction must be numerically robust. A hand-rolled
estimate_covariance(tol=1e-3)+GaussianSampler pipeline breaks on the wide panel (construction-fidelity-vs-model 0.5 โ invalid knockoffs). knockpy's KnockoffFilter end-to-end is the robust path (null FDR 0.0).
- The ~1/q detection floor is the ABSTAIN semantics. e-BH & knockoff+ cannot reject fewer than ~10 features at q=.10 โ the power plant must be dense (โฅ14) and single-feature probes (interaction, attribute) embedded in a dense background. This is exactly the Admit's "below power-contour โ ABSTAIN", confirmed.
Slice 4 (terminal) โ conservative-q calibration (PATTERN #2). The FDP over-shoot (~1.5ร nominal) on real dependent data is the same failure the analytic thresholds hit in #3/#4/#5/#7/#10/#19; the fix is the same: calibrate a conservative nominal qโฒ so the realised FDP โค .10 (the contour puts qโฒโ.05โ.06), report power at qโฒ and the IC-power contour as the published envelope, and re-confirm interaction-only (slice-2: RF-W 1.0 vs lasso-W 0.0) + attribute/ABSTAIN via KnockoffFilter. Operator-ruling flag candidate (like #0 PBO descope / #4 lag-tension): the knockoff power contour is intrinsically shifted right โ accept the @ICโ.04 operating point for the conditional-FDR class, or hold @IC=.03 with the ABSTAIN-below-contour envelope.