iter 7 — A2: the ξ/ρ low-power blind spot, hunted with distance correlation

BREAKS → DEFENDED → CLOSED 2026-07-10 · attack A2 (KB §1) · 496 real feature pairs × 4 markets · 14 blind-spot pairs found · advisory dCorr cross-screen added (KB-prescribed; no band) · clean 0/3

Claim under attack

ξ's detection boundary is n^−1/4 — rate-suboptimal on smooth alternatives (KB §1). Monotone smoothness is caught by the binding |ρ| gate, so the razor-edge claim was: "no real pair with strong smooth NON-monotone dependence sits below ALL binding pair gates (ξ < 0.50, |ρ| < 0.85) — the cascade never silently false-PASSes strong dependence."

Attack construction

Universe: 126 numeric non-label columns minus meta + raw price family (mirrors the declaration's collinearity exclusion) → 496 feature pairs.
Substrate: BTC/ETH/SOL/XRP @250, strided value-rows even-downsampled to n=800 (declaration methodology); usability filter per the probe.
Per pair per cell: hardened ξ (xi_max_pair, identity keys), |Spearman|, |r|², and distance correlation (Székely V-statistic, cached centered matrices — KB §1's prescribed higher-power comparator; null scale at n=800 ≈ 0.035).
Falsifier F1: dCorr ≥ 0.70 (~20× null — an attack evidence standard, not a band) in ≥2 markets, while ξ < 0.50 AND |ρ| < 0.85 in ALL markets.

ATTACK RESULT — the blind spot is real: 14 pairs

Worst examples (of 14)	max dCorr	max ξ	max \|ρ\|	strong cells
`bar_dispersion_entropy \| bar_bartels_rank_vn_ratio`	0.872	0.446	0.730	4/4
`bar_petrosian_fd \| bar_dispersion_entropy`	0.854	0.432	0.724	4/4
`bar_katz_fd \| bar_bartels_rank_vn_ratio`	0.843	0.435	0.757	4/4
`bar_petrosian_fd \| bar_katz_fd`	0.830	0.429	0.759	4/4
…10 more (full list in a2_results.json) — concentrated in the shipped bar_* complexity family (fractal dims, entropy, bartels, VG-clustering, sign-flux, L-kurtosis): strong smooth curved inter-relationships, exactly KB §1's predicted class.

DEFENSE — advisory dCorr cross-screen (KB §1's prescription, verbatim)

+ dcorr(x, y)                  # Székely V-statistic; even-downsample n=800; memory-safe
+ dcorr_advisory(...)          # xi_low_power_note = true when dCorr strong (≥0.70)
                               # while BOTH binding pair gates are quiet
+ aggregate_worst_cell         # per cell: dCorr on the vote-driving (argmax-ξ) pair;
                               # rollup: dcorr_worst + note in every candidate verdict

Advisory only — never a verdict. KB §1: "add a higher-power cross-screen (dCorr/HSIC or M-NN ξ) as an advisory corroborator — NOT a new banlist band (that would be meta-eval)." The 0.70 note line is the campaign's evidence standard, explicitly labeled "cut not banlist-authorized." Turning this class into a BAN requires an operator-declared, pre-registered dCorr calibration (07-DECLARATION playbook) — added to the operator handoff as H7.

Proof the defense closes the gap (v2 → oracle triage → v3)

Check	Result
12 blind-spot pairs silent on the full substrate	12/12 carry `xi_low_power_note = true` through the probe's own readout
2 pairs v2 flagged as "missed"	read ξ_worst = 0.501 on the FULL substrate → caught as WATCH by the frozen §B gate itself — not silent (v2's oracle demanded a note even where a binding gate had already fired; corrected in v3, honest-triage discipline)
Certified-orthogonal anchors (3 frozen negatives)	3/3 note-free — the advisory doesn't cry wolf
A5 substrate-validity regression · §B declaration self-test	SURVIVES 7/7 · PASS

Evidence

File	What
a2_blind_spot_attack.py / a2_results.json	v1 — 496-pair scan, the 14 breaching pairs, 42 strong-dCorr pairs total.
a2_blind_spot_attack_v2.py / a2_v2_results.json	v2 — probe-path advisory readouts for all 14 + 3 controls (blunt oracle; measurements canonical).
a2_blind_spot_attack_v3.py / a2_v3_results.json	v3 — corrected falsifier (silence AND no note); 0 breaches (final).
a2_stderr.log · a2_v2_stderr.log · a2_v3_stderr.log	Run logs under the 2 CPU / 2 GB scope.
regression_tests/test_a2_blind_spot.sh	Permanent regression test — refreshes measurements (v2), asserts corrected oracle (v3) SURVIVES.

Consequence

The last unattacked library item is closed. ξ's theoretically-predicted blind spot was found ALIVE in the shipped feature set — 14 real pairs — and is now permanently surfaced by a higher-power advisory cross-screen, while the meta-evaluation firewall stays intact (no new band; the ban decision is the operator's, handoff H7). Clean rounds: 0 of 3.

Open threads

evaluate_pooled positional tie-break default (advisory ξ CI path) — carried from iter 3; last remaining known thread.

iter 7 · attack A2 · verdict BREAKS → DEFENDED → CLOSED · one atomic commit: iter 7: A2 low-power blind spot — BREAKS→DEFENDED · append-only.