2026-06-06 · forex orthogonal candidate discovery · batch 3 + 4 · pane B (loop-batch34) · iteration 2

Iteration 2 — aggregate the 174-cell screen, cross-batch vs keepers, adjudicate

The full 174-cell-slice Spearman screen completed (exit 0). Rolled up new×new + new×143-prod, then ran the cross-batch new×keeper redundancy (20 new + 14 keeper kernels on the same windows, 4 assets × 3 slices). Adjudicated all 20 via the DISTRIBUTION rule. Verdict: 16 KEEP · 4 DROP.

174/174

cell-slices completed

KEEP (3 representatives · 2 watch)

DROP (3 collapsed · 1 location)

0/20

degenerate (value-health all pass)

▶ WHAT THE LOOP IS DOING NOW

Iteration 2 complete — dispositions set. This iteration surfaced new decisions (4 drops + 3 family collapses), so the loop-until-dry counter stays at 0/2.

⏭ NEXT PENDING TASKS

Iter 3: confirm the cross-batch verdicts on additional stress regimes (covid_crash · lowvol_grind · midcycle_range) — check no new ≥2-asset keeper duplicate emerges. If none → dry counter → 1.
Iter 4: final value-health + replication re-check; if no new decision → dry counter → 2 → STOP.
At dry (K=2): Attacker/Defender Challenge-and-Held, fix any conceded gap, then open the PR-from-worktree (PITSLARP body, no AI attribution, do not merge).

The 4 DROP decisions (with measured numbers)

DROP psd_spectral_centroid_normfreq → collapse to psd_wiener_spectral_flatness.
Intra-batch BAN: pair median 0.95, max 0.98, >0.95 on all 4 assets. The two spectral scalars are near-identical; keep the canonical Wiener-entropy representative (lower vs-prod max, 0.51 vs 0.58).

DROP porta_increment_sign_asymmetry → collapse to guzik.
Doubly redundant: >0.85 vs guzik on all 4 assets (pair med 0.78, max 0.96) AND >0.85 vs keeper mann_kendall_z on 3 assets (sign-fraction ≈ trend direction). Guzik is the magnitude-weighted (richer) representative.

DROP categorical_rqa_determinism_logret_sign → collapse to …diagonal_entropy.
Cross-batch duplicate: >0.85 vs keeper sign_two_state_entropy_rate replicated on 2 assets (XAG/XAU, worst 0.94) + 0.92 vs its RQA sibling. The diagonal-entropy member (single-asset overlap only) is the cleaner representative.

DROP ohlc_ordinal_pattern_index (location).
Persistent shadow of shipped order-flow/directional columns: vs-prod median 0.85, p90 0.89, >0.85 on all 4 assets across 105/174 cells (worst 0.94 vs ofi_skew_5k_bar). Never >0.95, but low independent dynamic range — informationless beyond the shipped surface.

Full adjudication — all 20

vs-prod = |ρ| vs the 143 shipped columns across all 174 cell-slices. vs-keeper = |ρ| vs the 14 batch-1-2 keepers across 4 assets × 3 slices. Breach = symbols where max|ρ|>0.85.

#	candidate	family	sub	prod med	prod max	prod breach	worst keeper (max)	keeper breach	disposition
96	sign_markov_entropy_production_3state	sign-Markov	signs	0.07	0.26	—	spectral_log_slope (0.22)	—	KEEP
119	palindromic_defect_sign_word	combinatorics-words	signs	0.13	0.33	—	sevcik_fd (0.27)	—	KEEP
25	pomeau_irreversibility_lag1	time-irreversibility	logret	0.13	0.39	—	mann_kendall_z (0.23)	—	KEEP
55	hvg_degree_assortativity	visibility-graph	close	0.17	0.30	—	hvg_clustering (0.58)	—	KEEP
118	critical_exponent_sign_word	combinatorics-words	signs	0.18	0.33	—	hvg_clustering (0.49)	—	KEEP
07	right_quantile_weight_tailweight	distributional tail	logret	0.16	0.56	—	anderson_darling_a2 (0.40)	—	KEEP
102	left_medcouple_lower_tailweight	distributional tail	logret	0.17	0.67	—	mann_kendall_z (0.60)	—	KEEP
63	hoeffding_d_midreturn_duration	copula-free non-mono	ret_dur★	0.18	0.49	—	anderson_darling_a2 (0.42)	—	KEEP
57	hvg_mean_shortest_path_length	visibility-graph	close	0.17	0.54	—	sevcik_fd (0.72)	—	KEEP
97	sign_markov_slem_mixing_3state	sign-Markov	signs	0.17	0.65	—	sevcik_fd (0.66)	—	KEEP
38	cumulative_residual_extropy	information	logret	0.27	0.61	—	mean_cosine_turning (0.80)	—	KEEP
28	guzik_increment_magnitude_asym	time-irreversibility	logret	0.23	0.63	—	mann_kendall_z (0.84)	—	KEEP·REP
70	categorical_rqa_diagonal_entropy	categorical-RQA	signs	0.13	0.45	—	sign_two_state_entropy (0.87)	XAU	KEEP·REP
72	psd_wiener_spectral_flatness	spectral	close	0.13	0.51	—	spectral_log_slope (0.58)	—	KEEP·REP
86	interquote_count_fano_factor	duration count-disp	qcount★	0.52	0.82	—	pietra_hoover_dur (0.65)	—	KEEP·WATCH
82	sign_negentropy_logcosh_gap	negentropy	logret	0.32	0.62	—	anderson_darling_a2 (0.87)	EUR	KEEP·WATCH
74	psd_spectral_centroid_normfreq	spectral	close	0.14	0.58	—	sevcik_fd (0.69)	—	DROP·collapse
27	porta_increment_sign_asym	time-irreversibility	logret	0.25	0.62	—	mann_kendall_z (0.87)	EUR·GBP·XAU	DROP·collapse
67	categorical_rqa_determinism	categorical-RQA	signs	0.13	0.47	—	sign_two_state_entropy (0.94)	XAG·XAU	DROP·collapse
44	ohlc_ordinal_pattern_index	candle ordinal	ohlc_bar	0.85	0.94	EUR·GBP·XAG·XAU (105c)	anderson_darling_a2 (0.04)	—	DROP·location

★ new substrate (ret_dur = logret×duration · qcount = per-bar quote_count). KEEP·REP = family representative (absorbs a collapsed sibling). KEEP·WATCH = clean on ≥2-asset replication but flagged for the downstream Chatterjee-ξ gate.

Intra-batch cand×cand (the binding correlations)

pair	median	max (cell)	>0.85 assets	>0.95 assets	verdict
psd_wiener_flatness ‖ psd_spectral_centroid	0.95	0.98 (XAU/gold_parabola)	all 4	all 4	BAN → collapse
guzik ‖ porta	0.78	0.96 (EUR/covid)	all 4	EUR	collapse
rqa_determinism ‖ rqa_diagonal_entropy	0.32	0.92 (XAU/gold_parabola)	XAU	—	single-asset → keep both*
sign_negentropy ‖ cumulative_residual_extropy	0.66	0.85 (XAG/gold_parabola)	XAG	—	single-asset → keep both

*rqa_determinism is nonetheless dropped — not for the single-asset sibling tie, but for its ≥2-asset cross-batch overlap with keeper sign_two_state_entropy_rate.

What iteration 2 did

Ran --mode aggregate over all 174 cell-slices → per-candidate worst+median |ρ| (cand×cand + cand×143-prod), written to AGGREGATE_VERDICT.json.
Ran a per-symbol breach analyzer → median / p90 / max + which assets breach >0.85/>0.95 + per-symbol value-health range.
Ran cross_batch_b34.py → rolled the 20 new + 14 keeper kernels on the same windows (4 assets × 3 slices) → new×keeper |ρ| with ≥2-asset replication, written to CROSSBATCH_VS_KEEPERS.json.
Adjudicated via the distribution rule → batch34_dispositions.json (16 KEEP, 4 DROP). All 20 pass the scale-aware value-health gate (min per-symbol range ≫ 1e-6; none point-mass).

Surviving family map (16 KEEP)

time-irreversibility → pomeau (25) + guzik (28, absorbs porta)
categorical-RQA → rqa_diagonal_entropy (70, absorbs determinism)
combinatorics-on-words → critical_exponent (118) + palindromic_defect (119)
sign-Markov → entropy_production (96) + SLEM (97)
spectral → wiener_flatness (72, absorbs centroid)
visibility-graph → mean_path (57) + assortativity (55)
copula-free non-monotone → hoeffding_d (63, ret_dur)
duration count-dispersion → fano_factor (86, qcount, watch)
negentropy / information → negentropy (82, watch) + CRExtropy (38)
distributional tail → right_quantile_weight (07) + left_medcouple (102)

Iteration 2 · generated 2026-06-06 · artifacts in worktree batch34_results/ (AGGREGATE_VERDICT.json · CROSSBATCH_VS_KEEPERS.json · batch34_dispositions.json). Next: iteration 3 — stress-regime cross-batch confirmation.