🎲 #2 block-permutation shuffled-label null GROUNDED

The first realness instrument: shuffle the answer in autocorrelation-preserving blocks; a genuine edge must survive, a spurious one must collapse to chance. It grounds cleanly on 870,304 real bars — correctly sized, detects real signal, and its careful block statistics are demonstrably load-bearing.

0.05

null test size (calibrated ∀L)

0.21

naive-iid over-rejection (block load-bearing)

FNR 1

U-shape blindness (published)

3 / 20

instruments grounded

Preflight (resource-only): load1 3.23 · 39 GiB available · si/so ~0 · ClickHouse active, readonly=2. Compute wall ~20 s, RSS ~270 MB (caps 5c/5G/no-swap, never hit). Runs on the grounded #0/#1 substrate.

In plain English

To decide whether a feature's link to the next move is real, we repeatedly shuffle the answers in chunks (keeping their rhythm intact) and ask: is the real link stronger than the shuffled ones? This instrument passes four honesty checks: (1) on a genuinely link-free case it wrongly "finds" a link only 5% of the time — exactly the tolerance we set; (2) a lazy shuffle that ignores the rhythm would falsely find links 21% of the time — proving the careful chunked shuffle is doing real work; (3) it does catch a guaranteed-real signal; and (4) it honestly reports its blind spot — a "smile-shaped" (U-shape) link is invisible to it, so those features get routed elsewhere.

Gate status — all pass

Gate	Result
Size calibration	test size 0.048 / 0.053 / 0.046 at L={25,50,100} — all in [.037,.063]	PASS
p-value uniformity	discrete chi-square uniform p 0.41 / 0.75 / 0.80	PASS
Necessity (naïve-iid)	block null 1.56× wider than iid → implied iid over-rejection 0.208 ≥ 0.20	PASS
Power (known-positive)	real look-ahead detected; power .43→.62→.75→.97; N_min 10,000	PASS
Known-duplicate	exact clone → identical IC & p-value	PASS
Envelope (U-shape)	even dependence marginal IC 3e-6, p 1.0 → FNR 1 (blind, expected)	PASS

Verify-before-report — six methodological corrections

This instrument is statistically intricate; each revision replaced an incorrect construction with the correct one (never tuned a threshold to pass): B_inner 59→99 (exact size 0.05) · M_outer 500→1500 (L=50 noise) · necessity → direct variance-inflation (the nested iid-size was construction-fragile) · U-shape → |zrank(feature)| (subsample-independent) · uniformity gate → discrete chi-square (continuous KS over-rejects the p-value grid discreteness at large M — only L=25 tripped: KS 0.005 but discrete-χ² 0.41). The discipline earned its keep here.

Envelope (published)

Detects: marginal-monotone rank associations (power .97 at N=10k).
Blind (FNR=1): zero-marginal-IC classes — U-shape shown (route to #11 CMI / #18 detector).
Interaction-only: no clean real exemplar under 03c → UNVERIFIABLE, route to #11.
Rare-veto: expected weak/blind → route to #18.

Verdict & what's next

GROUNDED — the first realness-axis instrument, running on the grounded substrate.

Next iteration → #3 effective-n deflation + BH-FDR (realness / multiple-testing): HAC long-run-variance effective-n + Benjamini-Yekutieli FDR, with the GARCH-heavy-tail necessity.