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

Iteration 1 — register 20 candidates, write clean-room kernels, smoke, launch the screen

The 20 newly-shortlisted parameterless candidates (batch 3 + batch 4) are registered in the multi-slice probe with clean-room numpy/scipy kernels, smoke-validated on real EURUSD@5 bars (0 degenerate, 0 null), and queued through the same Spearman multi-slice screen (174 cell-slices) that adjudicated batches 1-2.

candidates registered

20/20

kernels non-degenerate (smoke)

new substrates (ret_dur · qcount)

174

cell-slices queued

▶ WHAT THE LOOP IS DOING NOW

Running --mode full --batches 3,4 --max-hours 9 --out batch34_results inside the hard-capped forex-loops.slice (200% CPU / 4 GB, CPUWeight 10 → yields to production + the crypto loop). Read-only ClickHouse via the probe's polite ch() helper. Real bars only, computed in RAM; only result JSONs + this HTML are written.

⏭ NEXT PENDING TASKS

Iter 2 — aggregate: worst + median |ρ| for cand×cand and cand×143-prod across all 174 cell-slices.
Iter 2 — cross-batch redundancy: screen the new 20 against the batch-1-2 KEEP set (14 keepers) — not just among themselves.
Iter 2 — adjudicate via the DISTRIBUTION rule (median location · max tail · breach-mass · ≥2-symbol replication · scale-aware value-health) → KEEP / DROP / HOLD; drop intra- and cross-batch duplicates.
loop-until-dry (K=2): an iteration that surfaces no new decision increments the dry counter; stop at 2 + final summary page.

What iteration 1 did

Read all 20 candidate cards (findings/.../candidates/NN-<slug>.md) → extracted each estimator formula.
Wrote 20 clean-room numpy/scipy kernels + REGISTRY lines declaring each substrate + BATCHES[3], BATCHES[4]. GPL references (MZanin irreversibility lib, crqa, Sage, robustbase, Hmisc) were reimplemented FROM FORMULA; permissive refs (ts2vg MIT, statsmodels/networkx/scipy BSD, librosa ISC, ordpy MIT) used directly or mirrored. Provenance recorded per kernel.
Added 2 new substrates to prepare_substrates + roll_candidate: ret_dur = logret × bar-duration (for #63 Hoeffding-D) and qcount = per-bar quote_count (for #86 Fano factor — confirmed the column exists in fxview_cache.forex_bars).
Smoke-validated all 20 on EURUSD@5 live_tail (30,000 bars → 7,450 eval points, 125 legal prod cols): every kernel finite, 0 nulls, std>0, range>0 — none degenerate; parameterless by inspection (no free knob).
Launched the full 174-cell-slice screen in the slice with checkpoint/resume.

Smoke results — EURUSD@5 live_tail (real bars, measured)

Single-cell preview only — the binding numbers come from the 174-cell distribution in iter 2. topProd ρ = highest |ρ| vs a shipped column; worstCand ρ = highest |ρ| vs another new candidate.

#	candidate	substrate	std	range	topProd	ρ	worst cand-partner	ρ
25	pomeau_irreversibility_lag1	logret	0.163	2.11	ask_close	0.17	guzik	0.17
28	guzik_increment_magnitude_asymmetry	logret	0.043	0.28	lookback_twap_position	0.25	porta	0.79
67	categorical_rqa_determinism_logret_sign	signs	0.005	0.05	lookback_kaufman_er	0.06	sign_markov_slem	0.54
70	categorical_rqa_diagonal_entropy_logret_sign	signs	0.019	0.17	intra_intensity_qps	0.12	sign_markov_slem	0.43
118	critical_exponent_sign_word	signs	1.72	10	close_slope_rank_q	0.14	palindromic_defect	0.23
119	palindromic_defect_sign_word	signs	8.80	56	lookback_midprice	0.17	critical_exponent	0.23
96	sign_markov_entropy_production_rate_threestate	signs	0.002	0.04	lookback_kaufman_er	0.04	psd_wiener_flat	0.08
97	sign_markov_slem_spectral_gap_mixing_threestate	signs	0.041	0.20	lookback_garman_klass	0.08	rqa_determinism	0.54
72	psd_wiener_spectral_flatness	close	0.027	0.23	lookback_burstiness	0.12	psd_spectral_centroid	0.95
74	psd_spectral_centroid_normfreq	close	0.018	0.15	lookback_burstiness	0.16	psd_wiener_flatness	0.95
57	hvg_mean_shortest_path_length	close	3.40	20.9	lookback_burstiness	0.22	psd_spectral_centroid	0.58
55	hvg_degree_assortativity_coefficient	close	0.049	0.33	lookback_burstiness	0.20	hvg_mean_path	0.22
63	hoeffding_d_midreturn_duration	ret_dur★	0.005	0.05	session_fine_id	0.15	interquote_fano	0.20
44	ohlc_ordinal_pattern_index	ohlc_bar	4.34	14	bid_imbalance_ewma	0.85	critical_exponent	0.03
86	interquote_count_fano_factor	qcount★	18.9	138	lookback_intensity	0.51	hoeffding_d	0.20
82	sign_negentropy_logcosh_gap	logret	0.0005	0.005	lookback_permutation	0.38	cumulative_residual_extropy	0.65
38	cumulative_residual_extropy	logret	2.6e-5	2.3e-4	spread_high	0.19	sign_negentropy	0.65
07	right_quantile_weight_tailweight	logret	0.207	1.23	lookback_permutation	0.14	sign_negentropy	0.35
102	left_medcouple_lower_tailweight	logret	0.193	1.22	bid_close	0.18	porta	0.41
27	porta_increment_sign_asymmetry	logret	0.034	0.20	lookback_twap_position	0.25	guzik	0.79

★ new substrate. The single-cell preview already flags three intra-batch clusters to watch in the full screen (see below).

Predicted redundancy clusters (single-cell preview — confirm in iter 2)

Spectral pair (batch 3) — psd_wiener_spectral_flatness ↔ psd_spectral_centroid_normfreq = 0.95. Likely a collapse-to-representative; also watch both vs batch-1 keeper spectral_log_slope.

Logret sign/magnitude asymmetry — guzik (28) ↔ porta (27) = 0.79; pomeau (25) nearby. Three time-irreversibility items on the same logret substrate — expect at most one survivor.

Sign-stream structure — RQA determinism/entropy (67/70) ↔ sign-Markov SLEM (97) ≈ 0.43-0.54. Cross-batch: watch all four signs-items vs batch-2 keepers lziv_complexity_signs + sign_two_state_entropy_rate.

Candle ordinal vs order-flow — ohlc_ordinal_pattern_index ↔ bid_imbalance_ewma = 0.85 (vs a shipped column). Single cell — could be regime-specific; the distribution rule decides.

The 20 candidates — families & substrates

batch	#	candidate	family	substrate	kernel provenance
3	25	pomeau_irreversibility_lag1	time-irreversibility	logret	clean-room (MZanin GPL)
3	28	guzik_increment_magnitude_asymmetry	time-irreversibility	logret	clean-room
3	67	categorical_rqa_determinism_logret_sign	categorical-RQA	signs	clean-room (crqa GPL)
3	70	categorical_rqa_diagonal_entropy_logret_sign	categorical-RQA	signs	clean-room (crqa GPL)
3	118	critical_exponent_sign_word	combinatorics-on-words	signs	clean-room (Sage GPL; runs theorem)
3	119	palindromic_defect_sign_word	combinatorics-on-words	signs	clean-room Eertree (Sage GPL)
3	96	sign_markov_entropy_production_rate_threestate	sign-Markov	signs	clean-room (smoothed 3×3)
3	97	sign_markov_slem_spectral_gap_mixing_threestate	sign-Markov	signs	PyDTMC@MIT (reimpl)
3	72	psd_wiener_spectral_flatness	spectral	close	librosa@ISC (reimpl)
3	74	psd_spectral_centroid_normfreq	spectral	close	librosa@ISC (reimpl)
4	57	hvg_mean_shortest_path_length	visibility-graph	close	ts2vg@MIT + scipy@BSD
4	55	hvg_degree_assortativity_coefficient	visibility-graph	close	ts2vg@MIT + networkx@BSD
4	63	hoeffding_d_midreturn_duration	copula-free non-monotone	ret_dur★	clean-room (Hmisc GPL)
4	44	ohlc_ordinal_pattern_index	candle ordinal	ohlc_bar	ordpy@MIT (reimpl)
4	86	interquote_count_fano_factor	duration count-dispersion	qcount★	elephant@BSD (reimpl)
4	82	sign_negentropy_logcosh_gap	negentropy	logret	sklearn@BSD (reimpl)
4	38	cumulative_residual_extropy	information	logret	clean-room estimator
4	07	right_quantile_weight_tailweight	distributional tail	logret	clean-room (robustbase GPL)
4	102	left_medcouple_lower_tailweight	distributional tail	logret	statsmodels@BSD-3
4	27	porta_increment_sign_asymmetry	time-irreversibility	logret	clean-room

Cross-batch baseline — the 14 batch-1-2 KEEP set

Iter 2 tests the new 20 against these (in addition to among themselves). Same-substrate overlaps to watch: signs (lziv_complexity_signs, sign_two_state_entropy_rate), close-spectral (spectral_log_slope), visibility-graph (hvg_clustering, dhvg_indeg_outdeg_kld), durations (local_variation_lv, pietra_hoover_durations).

anderson_darling_a2 · dhvg_indeg_outdeg_kld · edge_spread_bps · hvg_clustering · kendall_taub_ret_spread · local_variation_lv · lziv_complexity_signs · mean_cosine_turning_angle · pietra_hoover_durations · sevcik_fd · sign_two_state_entropy_rate · spectral_log_slope · spread_sign_reversal_rate · mann_kendall_z (representative)

Method invariants (unchanged from batches 1-2)

Engine: Spearman rank→Pearson on a shared 200-bar de-gapped bar-level window; cand×cand + cand×143-prod, per cell-slice JSON, resumable.
Verdict canon: |ρ|>0.95 BAN · 0.85-0.95 WATCH · std/range~0 DEGENERATE; cross-symbol replication required.
Cells × slices: 7 EURUSD thresholds + 4 each for GBP/XAG/XAU, × 10 regime slices (covid, lowvol, parity, bank-stress, gold, midcycle, parabola, live-tail, asian, london-ny) = 174 cell-slices.
Data policy: real bars only · read-only ClickHouse · compute in RAM · no synthetic/surrogate series (CODEC-FOCI + IAAFT removed).
Resource: hard-capped forex-loops.slice 200% CPU / 4 GB, CPUWeight 10 — shared with the crypto loop; yields to production.

Iteration 1 of the loop-batch34 pane · generated 2026-06-06T04:30Z · next: iteration 2 (aggregate + cross-batch + adjudicate) once the 174-cell screen completes.