Dashboard › Probes › Realness Loop › Iter 29 · #14 operator-ruling

iteration 29 · 2026-07-24 · incrementality axis · operator-ruling checkpoint · laptop-drives-bigblack

📐 #14 Spanning intercept OPERATOR RULING

The full 7-gate battery is built and run at N_min=150k. Six gates ground robustly — coverage, both null-FPR gates, the spanned-collapse, the frozen-lag stability, and the omitted-premium Harden. The seventh, power@α=5bps, is ~0.78 (population, seed-averaged) — marginally below the frozen 0.8 bar — an intrinsic long-memory floor. The instrument is sound; the 5bps power shortfall is an operator call.

6 / 7

gates robust (4 seeds)

~0.78

power@5bps (bar 0.8)

α≈6bps

where power reaches 0.8

15 / 20

terminal (14 G · 1 R)

Preflight (resource-only): load1 5.13 · 39 GiB available · si/so ~0 (transient spike cleared) · ClickHouse active, readonly=2. 5c/5G/no-swap capped; another user's job used memory but no nasimubd heavy job.

The seven gates

Gate (§7 row 14)	Result (seed range)
Coverage & null-t	coverage 0.927–0.967 (≥.90)	PASS
Null-FPR (spanned feature)	0.0	PASS
Null-FPR across Hurst/t(3)/block	HAC & block-bootstrap both ≤.05 (natural + block-perm)	PASS
Substitution — B spanned collapses	\\|t\\| 0.2–0.35 (<1)	PASS
Frozen-NW-lag stability	t-medians [9.4,8.6,8.1,7.9], spread 0.10–0.23 (no lag cherry-picking)	PASS
Omitted-premium Harden	complete-basis \\|t\\| 0.7–1.2 spans · omit-premium t 7.8–10.4 spurious	PASS
Power @ α=5bps	population ~0.78 (seeds 0.81/0.76/0.79/0.75) · reaches 0.8 at α≈6bps	~0.78

Why the power sits at ~0.78 — the long-memory floor (verify-before-report)

Flat in N. Power is 0.72–0.85 (noisy) across N = 10k → 200k. The ~18–20% "bad-window" tail — extreme-long-memory windows whose HAC standard error blows up, giving t<3 even though the median t is ~9 — does not shrink with more data.
Scale-invariant. Tripling the plant's alpha leaves power and the t-median unchanged — magnitude can't help.
Block-bootstrap is worse. The §7 block-bootstrap t is more conservative (power 0.56 < HAC 0.72), so it controls the null-FPR under stress but cannot rescue power.

The power reaches 0.8 only at α≈6bps. The conservative t≥3 admit bar (not 1.96) also caps power and yields the null-FPR ~0. This is the intrinsic long-memory-effective-N floor — the same family as #4's HAC-effective-N and #12's power contour.

The ruling

(a) [recommended] accept the marginally-relaxed operating point (α≈6bps, or the published α-power / N-contour envelope) and GROUND #14 — the instrument is sound (6/7 robust incl. the Harden); the 5bps shortfall is the long-memory floor.
(b) hold the frozen power≥.8 @α=5bps bar → #14 does not graduate at 5bps (route weak-alpha features elsewhere; report the contour).
(c) require a more robust return construction (winsorised / rank-based strategy return) to shrink the bad-window tail — a larger build.

Precedent: #0 PBO descope, #4 lag-tension (later resolved by a pattern), #12 knockoffs (rejected). Per the #12 precedent, the next firing proceeds to #15 while #14 awaits the ruling (applied retroactively via a supersedes: row).