Finish · cross-candidate matrix

Cross-candidate orthogonality — CECP / statistical-complexity slate

Worst-cell rollup breaches the cap, but only in 3 of 209 cells (≤5%) — see the localization panel. Across the bulk of the grid the 3 CECP candidates are mutually orthogonal (medians well below cap); redundancy is confined to a few sparse, high-threshold cells where the ordinal histograms collapse.

LOCALIZED REDUNDANCY · 10/10 slices · 209 valid cells · generated 2026-06-10 05:07 UTC

0.942

worst |ρ| (any pair)

0.963

worst pairwise R²

0.970

worst leave-one-out R²

209

valid cells

Gate: cross-candidate orthogonality holds when max |ρ| ≤ 0.85, pairwise R² ≤ 0.85, and leave-one-out R² ≤ 0.85. Watch band 0.85–0.95; sibling/redundant above 0.95.

Localization — cells breaching the cap

3 of 209 valid cells (1.4%) exceed |ρ|>0.85, R²>0.85, or LOO R²>0.85. These drive the worst-cell rollup; the medians above show the typical case.

Cell	slice	bars	max \|ρ\|	max pairwise R²	max LOO R²
`BTCUSDT@750dbps`	`slice05` · 2025-01 → 02	642	0.635	0.963	0.970
`BTCUSDT@750dbps`	`slice01` · 2026-01 → 02	705	0.942	0.355	0.608
`BTCUSDT@750dbps`	`slice07` · 2024-07 → 08	885	0.610	0.705	0.861

Pairwise relationships (across all valid cells)

Pair	median \|ρ\|	max \|ρ\|	max pairwise R²	verdict
`statistical_complexity` ↔ `cecp_velocity`	0.175	0.758	0.963	SIBLING
`statistical_complexity` ↔ `binary_cecp`	0.280	0.942	0.624	WATCH
`cecp_velocity` ↔ `binary_cecp`	0.085	0.577	0.339	PASS

Leave-one-out R² (each candidate ~ joint of the other two)

How well each candidate is reconstructed from the other two together — the strictest within-slate redundancy test.

Candidate	Axis-1 vs panel	median LOO R²	max LOO R²	within-slate
🧭 Statistical Complexity (MPR) `bar_statistical_complexity`	BAN	0.188	0.970	REDUNDANT
🛰️ CECP velocity `bar_cecp_velocity`	PASS	0.050	0.969	REDUNDANT
🔳 Binary CECP `bar_binary_cecp`	WATCH	0.147	0.705	ORTHOGONAL

Per-slice worst case (regime spread)

Slice	window	valid cells	max \|ρ\|	max LOO R²
`slice01`	2026-01 → 02	21	0.942	0.615
`slice03`	2025-07 → 08	20	0.748	0.785
`slice05`	2025-01 → 02	21	0.635	0.970
`slice07`	2024-07 → 08	21	0.657	0.861
`slice09`	2024-01 → 02	21	0.743	0.668
`slice12`	2023-04 → 05	21	0.711	0.679
`slice14`	2022-10 → 11	21	0.654	0.662
`slice17`	2022-01 → 02	21	0.414	0.263
`slice20`	2021-04 → 05	21	0.640	0.563
`slice21`	2021-01 → 02	21	0.676	0.711

Method & scope

What this measures: candidate-vs-candidate redundancy within the CECP slate — orthogonal to the per-iteration pages, which measure each candidate vs the landed production panel only.
Slate run (deviation, documented): the contract Finish passes "PASS survivors", but only bar_cecp_velocity passed Axis-1; a single-candidate cross matrix is degenerate (no pairs; leave-one-out undefined). Per the Finish purpose ("catch siblings among the 3") and the never-block fork, the matrix was computed over all 3 candidates so the lone survivor can be checked for near-duplication against its WATCH/BAN CECP siblings.
Estimator: ordinal embedding m=3 / τ=1 (PR#513 ORDINAL_M/ORDINAL_TAU); cost-agnostic 100K-bar cap per cell so all candidates compute on the same rows; Spearman ρ + OLS R² (sklearn).
Read-only ClickHouse · 2-core@200% / 2 GB / spare-only cgroup.
Not a promotion decision. Promotion to production columns — and the open #20 caveat (confirm Binary-CECP sign-binarization is faithful to Pinto 2025, arXiv:2504.01974) — are operator/Terry decisions outside this loop.

Sources: results/cross_candidate_slice*.jsonl · registry orthogonal_candidates_v2.py.