iter 5 — A3: block-length misspecification of the CI-gated verdicts

SURVIVES 2026-07-10 · attack A3 (KB §3) · 18 regime-types × 7 pairs × 5 block lengths = 630 bootstrap CIs · 0 verdict flips · clean 1/3

Claim under attack (assumed true only to destroy it)

The BINDING worst-regime path CI-gates pooled |Spearman| with a stationary block bootstrap whose mean block length is hardcoded at 2√n. KB §3 flags block-length misspecification as a CI-distorting defect. The claim: this choice is not load-bearing — no ci_gated_redundant verdict on real pooled regime data changes across reasonable block lengths.

Falsifier

F1: any (regime-type, pair) whose CI-gated verdict differs anywhere across the 16× sweep {√n/2, √n, 2√n, 4√n, 8√n} (each floored at 10). One flip = the hardcoded choice silently decides frozen-line verdicts = claim BROKEN.

Mechanism fidelity

Pooling is aggregate_worst_regime-faithful: H4 one median full-tier threshold per symbol (16 cells), regime labels replicated verbatim (vol tercile × session × trend), per-symbol rank-normalization via the probe's own pool_rank_normalize, ≥40 rows/symbol/regime; n_pooled = 1,770–3,776 per regime.
CI machinery under test = the probe's own _boot_ci (B=400, α=0.10) + ci_gated_redundant (frozen 0.85/0.95): the block length is injected by patching the module-level sampler the probe calls; resampling = deterministic hash-derived stationary bootstrap over REAL pooled rows (no rng).
Statistic = |Spearman| — the CH-banlist binding gate.

Result — the 2√n choice is not load-bearing

Check	Result
Verdict flips across the sweep (126 regime×pair combos)	0
Bands represented in the combo set (non-vacuous test)	BAN, WATCH and PASS all present — none moved
Max CI-lower-bound spread across the 16× sweep	0.019 (high\|asian\|up · bar_dispersion_entropy\|trade_intensity) — an order of magnitude below the distance to any frozen line
Max single CI width (context)	0.160 (a small-n high-vol regime; still gated fail-safe by design)

Evidence

File	What
a3_block_length_attack.py	The attack — sweep + probe's own CI machinery; falsifier pre-declared in-file.
a3_results.json	Per-regime, per-pair, per-block-length CI bounds + verdicts (630 CIs).
a3_stderr.log	Run log under the 2 CPU / 2 GB scope.
regression_tests/test_a3_block_length.sh	Permanent regression test — re-runs the sweep, asserts SURVIVES.

Consequence

KB §3's misspecification concern is answered empirically for this substrate: verdict-level output is insensitive to a 16× block-length range, so the hardcoded 2√n stands (mechanism untouched — nothing needed defending). Noted for the record: arch (KB §3's ready Politis–White implementation) is not installed on this host; if this regression ever fails, a data-driven block length is the sanctioned fix and its absence becomes the KB-MISS to raise. Clean rounds: 1 of 3.

Open threads

A2 low-power blind spot · A4 cluster bootstrap (within-symbol vs pooled stitch) — unattacked.
evaluate_pooled positional tie-break default (advisory path) — carried.

iter 5 · attack A3 · verdict SURVIVES · one atomic commit: iter 5: A3 block-length sensitivity — SURVIVES · append-only.