⏱️ #1 future-perturbation invariance GROUNDED

The causality floor: a feature's value must not change when we corrupt the future. If it does, the feature is peeking ahead. This instrument checks that bit-for-bit (zero tolerance) — and it grounds cleanly on 870,304 real bars, catching a planted look-ahead every single time while a genuinely causal feature stays perfectly still.

V=0

causal feature (0 leaks / 221M checks)

1.000

power on planted look-ahead

1.0 → 0.0

warm-state FNR (cold-reset fixes it)

1 / 20

instruments grounded

Preflight (resource-only, per operator 2026-07-22): load1 0.26 (≤24) · 40 GiB free (≥8) · si/so 0 · ClickHouse active, readonly=2. Compute wall 15.1 s, peak RSS 220 MB (caps 5c/5G/no-swap, never hit). Slice 21 MB, read-only.

In plain English

Every honest predictor obeys one rule: today's number can only use today's and yesterday's data, never tomorrow's. This instrument is a lie-detector for that rule. It takes a feature, scribbles over all the future bars (deterministically — replaces them with a real value already seen, no invented numbers), recomputes, and checks whether any past value flinched. A flinch means the feature was secretly reading the future.

A well-behaved feature (an average of the last 32 bars) never flinched — 0 changes across 221 million checks. ✅
A cheating feature (one shifted to peek 3 bars ahead) flinched every time, and always at exactly the 3 positions it was peeking. ✅
The subtle trap: a feature that "normalizes by the whole series" is also cheating (the whole-series number includes the future). The detector misses it entirely if the engine is allowed to remember its earlier answer — and catches it perfectly once we force a clean restart. That's why "cold-reset the engine" is a hard requirement, not a nicety.

The four controls (technical)

Control	Real-data construction	Result
KNOWN-NULL (causal)	causal rolling mean L=32 vs constant-fill future corruption	FP = 0 / 221,187,473 bit checks · 0/512 decisions fired	PASS
KNOWN-POSITIVE (leaky)	real feature look-ahead-shifted k=3 (planted leak)	power = 1.000 (512/512); exactly k=3 positions changed each	PASS
KNOWN-DUPLICATE (subst.)	exact clone of the causal feature	V verdict identical (100% attribution)	PASS
HARDEN (warm-state)	normalize-by-full-series-std leak (occluded global dependency)	FNR 1.000 warm → 0.000 cold-reset + input-completeness	PASS

Why this isn't a vacuous pass

A bit-exact invariance test is only sound if the corrupted input set is provably complete and stateful engines are cold-reset (17b guard 2). The warm-state leg proves the failure it guards against, on real data: the same look-ahead leak is invisible (FNR = 1) to a warm engine that caches its statistic, and fully visible (FNR = 0) once cold-reset refits on the corrupted inputs. Cold-reset + input-set completeness are the difference between a sound instrument and a blind one — published in the detection envelope.

Detection envelope (published)

Detects: direct look-ahead (row-shift) · global-statistic / normalize-on-full-series leak (with cold-reset).
Blind without the Harden: warm-state / occluded global dependency (FNR = 1 unless cold-reset + input-completeness enforced).
Protected-class FNR (occluded / warm-state): with Harden 0.000 · without 1.000.
Zero tolerance: V(f)==0 is bit-exact — any single differing element = leak.

Verdict & what's next

GROUNDED — the causality floor is grounded. This is the first of the two substrate-gate instruments.

Next iteration → #0 CPCV purge/embargo (the leakage-free partition every out-of-sample number rides on). Substrate gate: evaluated only after #0 is decided — if either #1 or #0 fails to ground, the campaign HALTS before the realness/utility instruments. #1 clears its half.