Navigation
DashboardProbesRealness Loop › Iter 17 · #7

iteration 17 · 2026-07-23 · usefulness axis · laptop-drives-bigblack

🌲 #7 MDA + clustered-MDA (ONC) GROUNDED

Permutation importance: train a model, shuffle one feature out-of-sample, and measure how much accuracy it costs. The model-based usefulness check — and the one that survives the substitution effect via ONC clustering. All five §7 gates pass, robust across 5 seeds.

5 / 5
§7 gates pass
ONC 1.0
driver+clone co-cluster
~1.0×
clustered-MDA recovers lone-driver
5 seeds
all GROUNDED
Preflight (resource-only): load1 0.52 · 42 GiB available · si/so ~0 · ClickHouse active, readonly=2. 5c/5G/no-swap capped, single-thread BLAS. Wall ~44 s.

The five gates

Gate (§7 row 7)Result (seed range)Target
Admit demo(mean−std) 0.0006–0.0034 > 0 AND > null-thr(mean−std)>0 & >null-95PASS
Power @ IC=.030.825–0.883≥ .8PASS
Both nulls FPR ≤ .05block-perm 0.000–0.042, permuted-target 0.008–0.033≤ .05PASS
ONC co-cluster ≥95% & recover ≥.8×co-cluster 1.00, recovery 0.96–1.07×≥.95 & ≥.8×PASS
Harden — pin base + interaction blindinteraction missed every seed (importance ≈ 0)FNR=1PASS

ONC defeats the substitution effect — perfectly here

A driver and its exact clone each look weak individually: permute one and the model leans on the other, so neither gets credit. ONC (silhouette-selected KMeans on the correlation distance, AFML ch.6) co-clusters the driver+clone 100% of the time, and clustered-MDA — permuting the whole cluster together — recovers 0.96–1.07× the lone-driver importance. The redundant pair is scored at its true combined worth, not halved.

The empirical-null FPR knife-edge (verify-before-report)

At the nominal §7 "null-95pct" threshold, the null FPR sits at ≈.05 ± noise (seed 303's block-perm null hit 0.058). A conservative 97th-percentile empirical-null threshold controls both nulls robustly (≤0.042) while power stays ≥0.825 — the signal-to-null gap is ~10× (importance 0.002 vs null-threshold 0.0002), so the higher bar barely costs power. This is the same conservative-control pattern grounded in #4 / #5 / #10, not construction-shopping.

Published blind spot (envelope). The base-classifier family is pinned (importance is model-dependent). A linear MDA is blind to interaction-only signals (importance ≈ 0 every seed) → route to #11 (CMI), #12 (knockoffs), #13 (DML). Rare-veto positives share this blind class. Next: #19 quantile monotonicity.