📉 #5 IC-decay / half-life GROUNDED

Does the signal peak at a tradable forward horizon and decay monotonically (healthy) rather than spike or stay flat (snooping / leakage)? All four §7 gates pass, robust across 5 seeds.

4 / 4

§7 gates pass

ρ = −1.0

monotone post-peak decay

2–3×

OLS FWER vs HAC (HAC needed)

5 seeds

all GROUNDED

Preflight (resource-only): load1 0.38 · 42 GiB available · si/so ~0 · ClickHouse active, readonly=2. 5c/5G/no-swap capped, single-thread BLAS.

The four gates

Gate (§7 row 5)	Result (seed range)	Target
Admit demo	h=3, peak\\|IC\\| 0.53, HAC-t ~78, H½ 10, post-peak ρ −1.00*	peak≥.03, t≥3, H½≥2, ρ≤−.9	PASS
Power	1.00	≥ .8	PASS
Both nulls FWER ≤ .05 & OLS exceeds	HAC block-perm 0.020–0.033, circshift 0.010–0.045; OLS 0.09–0.15 > HAC 0.04–0.05	≤.05 & OLS exceeds	PASS
Harden — flat-curve leakage flag	far-horizon + flat-blend both flagged; healthy positive passes	leaks flagged	PASS

The load-bearing idea: overlapping horizon returns make HAC necessary

Unlike #4 (a single 3s return, near-white → HAC not load-bearing at the bar level), #5's horizon returns ret_h[i]=log(close[i+h]/close[i]) overlap for h≥2 (adjacent bars share h−1 terms — the #0 straddling mechanic). So the rank-IC contribution is autocorrelated, and the Newey-West correction genuinely matters: on the autocorrelation-preserving (circular-shift) null the naïve OLS-SE family-wise-errors at 0.09–0.15 while HAC controls at 0.04–0.05 — ~2–3× worse without HAC.

Four construction errors caught (verify-before-report)

The KNOWN-NULL must permute the FEATURE, not the returns. Permuting each ret_h independently destroyed the returns' cross-horizon correlation → 7 independent tests → BH FWER inflated (0.068). Block-permuting the feature keeps the overlap correlation (positively-correlated horizons → BH conservative).
The leak archetype is a far-horizon look-ahead, not a mean-of-returns. The clean leaks are ret_20 (peaks at the last horizon → no decay observed → flagged) and a mid+far blend (fails the decay test); both flagged, the healthy positive is not.
The OLS-necessity threshold is "exceeds .05", not an arbitrary floor. 0.09 > 0.05 with HAC at 0.04 satisfies "OLS-SE must exceed".
The BH-FWER heavy-tail knife-edge → the #10/#4 permutation family threshold. Analytic BH on N(0,1) p-values is anti-conservative on the heavy-tailed overlapping-return HAC-t null (0.06 on one seed). A permutation-calibrated Westfall-Young max-t threshold at a conservative 97th percentile controls the family-wise FPR by construction.

Published blind spot (envelope). A flat / no-decay IC curve is a leakage/snooping signature — flagged, not passed. Non-monotone real signals are also flagged (FNR=1) → route to #7 (MDA), #19 (quantile monotonicity). Next: #7 MDA.