Dashboard › Probes › Realness Loop › Iter 25 · #12 operator-ruling

iteration 25 · 2026-07-24 · conditional axis · operator-ruling checkpoint · laptop-drives-bigblack

🎛️ #12 Model-X Knockoffs OPERATOR RULING

Slice 4 tested the two remaining resolutions to the shifted power contour — and both are falsified. The standard Gaussian model-X knockoff cannot hit power≥.8 @IC=.03 (ceiling 0.67 across four statistics) and delivers an irreducible realised FDR ≈ 1.3× nominal (conservative-q calibration does not work). Both frozen §7 @IC=.03 gates are unmet with the standard construction — this is now genuinely the operator's call.

0.67

power ceiling @IC=.03 (need .8)

~1.3×

realised FDR (irreducible)

✓

interaction / attribute grounded

ruling

TSKI / envelope / route→#13

Preflight (resource-only): load1 2.70 · 36 GiB available · si/so ~0 · ClickHouse active, readonly=2; a concurrent nasimubd DOGEUSDT research job ran at ~1 core (spare-only, no contention). 5c/5G/no-swap capped.

Falsification 1 — power ceiling @IC=.03 (four statistics)

knockoff statistic	power @IC=.03	FDP
lasso (LCD)	0.67	0.148
mlr (masked likelihood ratio)	0.446	0.126
ridge	0.295	0.052
lcd	0.455	0.081

No valid statistic reaches power ≥ .8 @IC=.03. The conditional-FDR machinery is intrinsically too conservative for the weakest IC=.03 (t≈3) signal — the marginal #4 rank-IC gets power 1.0 there precisely because it tests a weaker hypothesis.

Falsification 2 — conservative-q calibration (PATTERN #2) does not work

operating point	nominal q=.10	q=.07	q=.05
IC=.05 (power / FDP)	0.96 / 0.126	0.60 / 0.143	0.0 / 0.0
IC=.06 (power / FDP)	0.98 / 0.135	0.90 / 0.131	0.10 / 0.039

Tightening the nominal q does not reduce the realised FDP — it stays stuck at ~0.13 until the ~1/q detection floor kills the filter (power → 0). There is no operating point with FDR≤.10 AND power: the discrete floor blocks a smooth q-vs-FDP tradeoff, and the leaking nulls carry signal-like W (the residual exchangeability violation from the long-memory + near-singular real panel) that no threshold separates. The FDP failure is irreducible by calibration — unlike #3/#4/#5/#7/#10/#19 where the conservative-threshold pattern worked.

What IS grounded (unaffected)

Interaction-only non-linear W (slice 2): RF-W 1.0 vs lasso-W 0.0 — the knockoff-vs-rank-IC/MDA differentiator.
Attribute / ABSTAIN (slice 2): signal-carrier A ≫ decoy B; exact clone → MVR s=0 → neither.
Null-only FDR 0.0; power 0.96–0.98 at IC≥.05 — the instrument detects; only the FDR is stuck at ~1.3×.

The residual ~1.3× is the Harden's "iid-exchangeability = the #1 gaming" as a residual effect — slice 2 correctly falsified it as the cause of the 0.25 plain-knockoff inflation; the residual ~1.3× is the genuine autocorrelation cost that survives the correct knockoff+/e-BH rule.

The ruling

Standard Gaussian model-X knockoffs deliver realised FDR ≈ .13 (not ≤.10) with power ≥ .8 only at IC ≥ .05 on the real panel, and cannot be q-calibrated to FDR ≤ .10 without dying. The untried path is a proper TSKI (HMM / deep generative knockoff modelling the long-memory joint distribution) — a substantial build; the naive whitenings (AR, uniform fracdiff, thinning, spacing) were all falsified in slices 2–3.

(a) commission a proper TSKI / time-series knockoff — major construction effort.
(b) accept the realised-FDR≈.13 + IC≥.05 power-contour envelope (relax the frozen @IC=.03/FDR≤.10 gate for the conditional-FDR class).
(c) [recommended] route the conditional-axis FDR-selection role to #13 DML and record #12 as construction-limited on this real long-memory panel — the conditional axis stays covered by #11 GROUNDED (interaction-only) + #13 (next).

Precedent: operator-ruling flags at #0 (PBO descope, 2026-07-22) and #4 (lag-tension, iter-14). Per recommendation (c), the next firing proceeds to #13 DML while #12 awaits the ruling; the ruling applies retroactively via a supersedes: row.