Dashboard › Probes › Realness Loop › Iter 22 · #12

iteration 22 · 2026-07-23 · conditional axis · bounded slice 1 · laptop-drives-bigblack

🎛️ #12 Model-X Knockoffs CHECKPOINT

The FDR-controlled conditional-independence test that — unlike CMI (#11) — does not require a perfectly-capturing conditioning set. Slice 1 builds the construction (MVR s-solver + Ledoit–Wolf shrinkage), and two verify-before-report findings emerge: the real microstructure panel is near-singular so the naïve equicorrelated knockoff is a dead filter (MVR fixes it → power 0.95), and iid knockoffs on autocorrelated bars inflate FDR — the Harden's "#1 gaming", demonstrated → slice-2 needs block/TSKI.

0.95

power (dense plant, MVR)

Π 1.0

derandomization stable

0.25

null-FDR (iid → inflated)

slice 1

block/TSKI next

Preflight (resource-only): load1 2.07 · 38 GiB available · si/so ~0 · ClickHouse active, readonly=2. 5c/5G/no-swap capped; single-thread BLAS. Panel staged read-only at /tmp/i12_slice.parquet (BTCUSDT@250, 8 full-spread real features).

Slice-1 gate status

Gate (§7 row 12, slice 1)	Result (seed 20260723)	Target
Power (dense main-effect plant)	0.95 (10-signal, achieved IC 0.038)	≥ .8	PASS
Derandomization Π	1.0 (median signal Π)	≥ .5	PASS
FDP (with power)	0.126	≤ .10	iid-inflated
Null FDR (block-perm target)	0.25 any-selection	≤ .10	iid-inflated
Exchangeability 2nd-moment	0.0267	≤ .02	borderline

Finding 1 — the near-singular real panel (MVR is not optional)

The empirical feature covariance is near-singular: ofi ≡ turnover_imbalance (ρ = 1.00), plus price_impact/volume_per_trade (−0.71) and ofi/aggression_ratio (0.78) → λ_min 0.0012, cond 2211. The naïve equicorrelated construction sets one scalar s = 2λ_min ≈ 0 for all features → knockoffs are near-copies → the filter is dead (power 0).

MVR + Ledoit–Wolf fixes it by allocating per-feature s: the duplicate pair gets s ≈ 0 (knockoff = copy → it can never be spuriously attributed — the correct ABSTAIN behaviour, and the ready ρ=1 exemplar for slice-2), while independent signals get s ≈ 1 → full power. Result: power 0.95 on a dense 10-signal plant, derandomization Π 1.0. (A single planted signal was also diagnosed to hit the knockoff+ 1/q detection floor — you cannot reject < 1/q features — so the power plant is dense per the spec, with plain-knockoff selection.)

Finding 2 — iid-exchangeability is the #1 gaming, confirmed on real data

The real features carry row-autocorrelation (max lag-1 AC 0.36, price_impact). Model-X Gaussian knockoffs sample each row as an iid draw from N(0,Σ) — which the consecutive-bar autocorrelation violates → the null features' statistics are not properly symmetrised → FDR leaks: null-FDR 0.25, FDP 0.126 (both > .10). This is precisely the Harden's warning — "iid-exchangeability = the #1 gaming" — surfaced by the slice-1 null gate rather than assumed. The borderline exchangeability (0.0267) is the near-singular panel + shrinkage residual.

Slice 2 (next firing) resolves it: block / TSKI knockoffs (a time-series construction that respects the row-autocorrelation) to restore FDR ≤ .10 ∀H ≤ .85; exchangeability on a conditioned panel; the interaction-only Harden plant (non-linear W statistic — knockoffs must still select what rank-IC/MDA miss); and attribute @ ρ≤.8 / ABSTAIN @ ρ≥.9 (the ofi≡turnover ρ=1 pair, already s≈0, is the exemplar).