🔬 SEAL — control-substrate self-validation SEAL-PASS

Before any instrument may judge a feature, the ruler itself must be checked against cases whose answer we already know. This iteration builds the four real-data-only controls and the shared IC + HAC detector, runs them on 870,304 real BTCUSDT bars, and confirms all four behave correctly. The lab is open.

4 / 4

controls validated

870,304

real bars (BTC@250, 3s)

2.34×

naive-SE understatement (HAC needed)

0 / 20

instruments grounded (unchanged)

Preflight (resource-only, per operator 2026-07-22): load1 0.38 (≤24) · 38 GiB free (≥8) · si/so 0 · ClickHouse active, readonly=2 accepted. Sidecar & kintsugi are intentionally stopped — excluded from preflight by operator instruction. Compute wall 17.6 s, peak RSS 364 MB (caps 5c/5G/no-swap, never hit).

In plain English

A metric that claims "this feature is real" is only trustworthy if we've first shown it gets the known cases right. SEAL builds four such known cases entirely out of real market data (no invented numbers, per the no-synthetic-data rule) and checks the detector against each:

A guaranteed-real signal (let the "feature" secretly peek one bar into the future) → the detector must shout. It did (t-score 43.8). ✅
A guaranteed-nothing case (shuffle the answer in blocks so any link is destroyed) → the detector must stay quiet. It fired on only 0.2% of 500 shuffles — quiet. ✅
An exact copy of a feature → the redundancy check must read "identical." Correlation came out exactly 1.000000. ✅
How much data it needs → by shrinking the sample we found it needs ~100,000 bars to reliably catch a real-but-weak effect. ✅

It also confirmed why the careful (HAC) statistics matter: the naïve textbook formula understates the real uncertainty by 2.34× here — using it would manufacture false confidence. That exact figure was predicted in advance by an earlier probe (P1, ~2.3×), so it's a check we could score before running.

The four controls (technical)

Control	Real-data construction	Result	Gate
KNOWN-DUPLICATE	exact clone of `vwap_close_deviation`	Spearman ρ = 1.000000; ΔIC = 0 (IC −0.31507 both)	ρ = 1 exactly	PASS
KNOWN-POSITIVE	`feature[t] = r[t+1]` (return look-ahead 1 bar)	IC = ρ₁ = 0.132, HAC-t = 43.8 (naive-t 123.1); fires	fires, \\|HAC-t\\|≥3	PASS
KNOWN-NULL	`vwap_close_deviation` × block-permuted target (L=100, 500 perms)	HAC FWER = 0.002 (silent; max\\|t\\| 3.64)	FWER ≤ 0.05	PASS
POWER-vs-N	contiguous subsample of KP signal, lag k*=29 (ρ=0.064)	power .09→.12→.29→.66→.84 ⇒ N_min = 100,000	N_min derivable	PASS

HAC load-bearing (necessity): naive-iid SE understates the true SE by 2.34× on the real aligned measurement (P1-anchored ~2.3×) → naive manufactures significance; gate ≥1.5 ✅. Dialed-IC N_min calibration → UNVERIFIABLE-PARK (untraceable without generating data; only the achievable-effect N_min is reported).

Verify-before-report — the necessity-leg correction trail

The four core controls passed on every run; the first two runs returned SEAL-FAIL on the necessity sub-gate only, each a harness mis-specification caught before trusting the number:

Naive FWER on the block-permutation null — that null destroys the feature↔target alignment, so its IC variance collapses to ≈1/N and naive SE is coincidentally correct (naive = HAC = 0.002). Wrong vehicle.
Naive FWER on a circular-shift null — also naive = HAC = 0.008. The aligned product's persistence is signal-driven (IC −0.315), not marginal; a signal-free null isn't wide (§4: "the null lacks that structure").
Deterministic real-aligned SE-understatement factor (adopted) = 2.34×, pre-anchored by P1 — the correct, feature-invariant measure. The circular-shift FWER stays in the JSON as the null-structure diagnostic.

Not goalpost-moving: attempts 1–2 measured naive-inflation on nulls where it provably cannot manifest for this feature; attempt 3 measures the intended concept with a statistic committed in advance.

Verdict & what's next

SEAL-PASS — the four real-data controls and the Spearman-IC + Newey-West-HAC detector primitive are validated ground truth. The lab is open. The graduation ledger stays 0 GROUNDED (SEAL grades none of the 20).

Next iteration → #1 future-perturbation invariance (causality floor): corrupt post-decision bars in a read-only slice and require the verdict be bit-identical (V(f)==0). Then #0 CPCV purge/embargo. Substrate gate: both must ground or the campaign HALTS before the realness/utility instruments.