2026-06-06 · forex orthogonal candidate discovery · batch 3 + 4 · pane B (loop-batch34) · iteration 6 · FINAL

Iteration 6 — FINAL: loop-until-dry STOP + Challenge-and-Held

The final independent audit (value-health · parameterless-by-inspection · intra-batch among survivors) surfaced no new decision ⇒ second consecutive dry iteration ⇒ loop-until-dry K=2 → STOP. Then an adversarial Attacker/Defender review of the verdict: 9 HELD, 3 CONCEDED→FIXED, 0 verdict flips. Final, locked: 15 KEEP · 4 DROP · 1 HOLD of the 20 batch-3/4 candidates.

2 / 2

loop-until-dry → STOP

KEEP (3 reps · 2 watch)

4 / 1

DROP / HOLD

9 / 3

HELD / CONCEDED (0 flips)

■ STOP — the loop is dry

Iter 5 (pooled-union cross-batch) and iter 6 (final audit) both surfaced no new decision (K=2). Iteration 6 confirmed: (a) value-health — 0 of 15 survivors near-point-mass (smallest min per-symbol median range 4.3e-4 ≫ the 1e-9 degeneracy floor); (b) parameterless — all 15 kernels knob-free (only declared structural-substrate conventions: 200-bar lookback, ODB bar, l_min=2, FFT bins by sample length, native count grid); (c) intra-batch among survivors — highest pair |ρ| = 0.69 (guzik|left_medcouple), zero pairs >0.85 on any asset.

Final verdict — 15 KEEP

#	candidate	family	substrate	disposition
25	pomeau_irreversibility_lag1	time-irreversibility	logret	KEEP
28	guzik_increment_magnitude_asymmetry	time-irreversibility	logret	KEEP·REP·WATCH
118	critical_exponent_sign_word	combinatorics-on-words	signs	KEEP
119	palindromic_defect_sign_word	combinatorics-on-words	signs	KEEP
96	sign_markov_entropy_production_3state	sign-Markov	signs	KEEP
97	sign_markov_slem_mixing_3state	sign-Markov	signs	KEEP
70	categorical_rqa_diagonal_entropy_logret_sign	categorical-RQA	signs	KEEP·REP·WATCH
72	psd_wiener_spectral_flatness	spectral	close	KEEP·REP
57	hvg_mean_shortest_path_length	visibility-graph	close	KEEP
55	hvg_degree_assortativity_coefficient	visibility-graph	close	KEEP
63	hoeffding_d_midreturn_duration	copula-free non-monotone	ret_dur★	KEEP
86	interquote_count_fano_factor	duration count-dispersion	qcount★	KEEP·WATCH
38	cumulative_residual_extropy	information	logret	KEEP
07	right_quantile_weight_tailweight	distributional tail	logret	KEEP
102	left_medcouple_lower_tailweight	distributional tail	logret	KEEP

4 DROP + 1 HOLD (not shortlisted)

#	candidate	disposition	basis (pooled-union denominator)
74	psd_spectral_centroid_normfreq	DROP·collapse	med 0.95 / max 0.98 vs flatness, >0.95 all 4 assets → flatness
27	porta_increment_sign_asymmetry	DROP·collapse	vs guzik all 4 + vs mann_kendall_z {EUR,GBP,XAU} → guzik
67	categorical_rqa_determinism_logret_sign	DROP·collapse	vs sign_two_state_entropy {GBP,XAG,XAU}, worst 0.98 → rqa_diagonal_entropy
82	sign_negentropy_logcosh_gap	DROP·connector	vs anderson_darling_a2 {EUR,XAG,XAU} pooled, worst 0.90
44	ohlc_ordinal_pattern_index	HOLD·location	vs-prod med 0.85, all 4 assets, 105/174 cells, never >0.95, diffuse 2-col shadow

Challenge and Held

Adversarial Attacker/Defender review (both read-only, in-slice) of the 15/5 verdict. Each challenge → HELD with evidence, or CONCEDED → FIXED. Net: 9 HELD, 3 CONCEDED→FIXED, 0 verdict flips.

⚒ 3 CONCEDED → FIXED (applied before this page)

guzik WATCH — guzik’s worst keeper mann_kendall_z = 0.84 sits 1 bp under the 0.85 breach line. KEEP held (0 breach assets), but disposition relabeled KEEP_REPRESENTATIVE → KEEP_REPRESENTATIVE_WATCH and flagged for the ξ gate.
Single breach denominator — the dispositions JSON exposed three breach fields (within-grid / pooled / by-grid). Collapsed to ONE verdict-bearing keeper_breach_symbols_pooled (pooled-union across all 10 slices, the batch-1-2 precedent); the others demoted to x-provenance diagnostics. Recomputing under the single denominator changes no verdict.
ohlc_ordinal DROP → HOLD — relabeled DROP_LOCATION → HOLD_LOCATION to match the batch-1-2 candle_shoelace_area precedent (persistent median shadow, never >0.95, diffuse across ≥2 prod cols). Same practical outcome (held out of the shortlist), more precedent-consistent vocabulary.

🛡 9 HELD (with evidence)

1 · negentropy DROP “regime-cherry-picked”? HELD. The denominator is the pooled-union across all 10 slices (batch-1-2 precedent). negentropy vs anderson breaches on 3 distinct assets {EUR,XAG,XAU}; the by-grid split (EUR-calm + metals-stress) is just where the redundancy manifests — 3 assets is not a single-asset artifact. The strict within-slice alternative is not the established method.

2 · rqa_diagonal KEEP vs determinism DROP “noise”? HELD. Under the same pooled denominator the decisive metric is breach-asset-count, not max|ρ|: determinism {GBP,XAG,XAU}=3 → DROP; diagonal {XAU}=1 → KEEP. The 0.96-vs-0.94 tail is irrelevant; the clique correctly collapses to one representative.

3 · ohlc DROP false (never >0.95)? PARTIALLY CONCEDED (→ relabeled HOLD, see fixes) / HELD on not-shortlisting: median 0.852 with all-4-asset breach across 105/174 cells is a location verdict; never >0.95 means it is HOLD_LOCATION, not KEEP — its median is higher than every KEPT candidate (≤0.42) and higher than the batch-1-2 HOLD example (0.72).

4 · fano KEEP@0.52 vs ohlc DROP@0.85 “double standard”? HELD. Both judged on the same median-location axis. fano 0.52 with 0 cells >0.85 passes; ohlc 0.852 with all-4-asset breach fails. 0.52 vs 0.85 is a real magnitude gap, not opposite-half cherry-picking. (fano stays KEEP·WATCH — flagged precisely because 0.52 is the highest surviving median.)

5 · Fano hidden threshold knob? HELD. The counting window is the ODB bar — project-wide structural substrate (like the 200-bar lookback), not a per-feature analyst knob. Card 86 declares the count pinned to native per-bar granularity. The screen tests across all 7 thresholds per cell. Var/mean has no tunable parameter.

6 · RQA l_min=2 is a free knob? HELD. l_min=2 is the irreducible definitional floor (a length-1 diagonal is a point, not a line) — the parameterless-by-convention value declared on cards 67/70. Marwan’s l_min∈[2,10] sweeps are the free-knob variant this candidate explicitly does not take.

7 · guzik needs WATCH CONCEDED → FIXED (see above).

8 · spectral FFT bins lookback-induced? HELD. K is a deterministic function of the 200-bar structural substrate, declared a non-knob on cards 72/74. flatness vs spectral_log_slope keeper = 0.58 (0 breach) — no inflation; centroid is the sibling that collapses.

9 · CRExtropy degenerate / lenient? HELD. min per-symbol median range 4.3e-4 = 4.3×10⁵ × the 1e-9 degeneracy floor → not near-point-mass; the rank-based screen is scale-invariant. CRExtropy has 0 pooled keeper breach (worst 0.80); its negentropy tie is single-asset (XAG) — kept under the identical ≥2-asset rule that kept rqa_diagonal’s XAU-only.

10 · three denominators CONCEDED → FIXED (single pooled denominator).

11 · ohlc names two prod cols CONCEDED→reframed: a diffuse 2-col block shadow is the signature of a location HOLD (low-rank shadow of the directional subspace), not a single-column collapse — supports HOLD, applied.

Surviving family map (15 KEEP)

⏳ Time-irreversibility — 25 pomeau · 28 guzik·REP·WATCH

♻️ Categorical-RQA — 70 diagonal-entropy·REP·WATCH

🔤 Combinatorics-on-words — 118 critical-exp · 119 palindromic-defect

🔀 Sign-Markov — 96 entropy-prod · 97 SLEM

📡 Spectral — 72 wiener-flatness·REP

🕸️ Visibility-graph — 57 mean-path · 55 assortativity

🔗 Copula-free non-monotone — 63 Hoeffding-D (ret_dur★)

⏱️ Duration count-dispersion — 86 Fano·WATCH (qcount★)

🔢 Information — 38 CRExtropy

📐 Distributional tail — 07 right-quantile · 102 left-medcouple

Artifacts

Worktree findings/.../batch34_results/: 174 per-cell-slice JSONs · AGGREGATE_VERDICT.json · CROSSBATCH_VS_KEEPERS_iter{3,4}_*.json · batch34_dispositions.json (final, with Challenge-and-Held fixes).
Probe: 20 clean-room kernels + 2 new substrates (ret_dur, qcount) in multislice_orthogonality_probe.py (GPL refs reimplemented from formula).
Two new substrates were required: #63 Hoeffding-D (ret_dur = logret×duration) and #86 Fano (qcount = per-bar quote_count).

Iteration 6 (FINAL) · generated 2026-06-06 · loop-until-dry K=2 reached · PR-from-worktree (loop-batch34) opened for review with this Challenge-and-Held in the body · not merged.