CODEC / FOCI orthogonality cascade — what's grounded

Exploration Gate · clean leakage-free run · read-only ClickHouse · every number traces to iterations.jsonl

✓ EXPLORATION GATE PASSED. Canary ofi↔turnover_imbalance flagged in 437/437 dense cells (100%); Type-I = 0. A gate pass means proceed to Selection, not adopt.

437cells100%canary0.881walk-forward Jaccard (clean)0.839non-canonical Jaccard0.0Type-I

Walk-forward — leakage-free (discover-on-train / validate-on-held-out)

325 full-tier cells · mean Jaccard 0.881 (min 0.2) · stable (≥0.20) in 325/325 · PBO 0.0 · Type-I 0.0 · Type-II 0.052. Purge 200, embargo 10, fixed-width rolling train.

Discriminating cut (audit fix): the flagged set is dominated by ρ≈1 structural duplicates that persist under any split. Excluding them, the non-canonical mean Jaccard = 0.839 — the honest measure of whether the cascade's discovered redundancies generalize.

regime	cells	mean Jaccard	min
bear_2022	16	0.952	0.75
bull_2021	64	0.868	0.587
bull_2024	64	0.883	0.482
current_2026	15	0.897	0.673
range_2025	59	0.903	0.611
recovery_2023	59	0.852	0.494
stress_2022	48	0.875	0.503

by threshold	100 dbps	250 dbps	500 dbps
mean Jaccard	0.967	0.903	0.762

Spot-checks (traceable to real records): BTCUSDT@100/slice14 J=0.5779 (14 folds) · BTCUSDT@250/slice14 J=0.7613 (5 folds) · BTCUSDT@500/slice14 J=0.6304 (5 folds)

Exploration Gate — agreement with known verdicts

pair	prior	agree	rate
`ofi↔turnover_imbalance`	REDUNDANT	437/437	100%
`agg_record_count↔individual_trade_count`	REDUNDANT	428/437	97%
`vwap↔close`	REDUNDANT	437/437	100%
`kyle_lambda_proxy↔price_impact`	ORTHOGONAL	437/437	100%
`trade_intensity↔aggression_ratio`	ORTHOGONAL	437/437	100%
`vwap_close_deviation↔aggregation_density`	ORTHOGONAL	437/437	100%
`kyle_lambda_proxy↔duration_us`	ORTHOGONAL	437/437	100%

Held disagreement(s): agg_record_count↔individual_trade_count in 9 cells (threshold-sensitive, ξ straddles 0.80; carried to Selection).

Canary by symbol / threshold

symbol	ADAUSDT	BNBUSDT	BTCUSDT	ETHUSDT	LTCUSDT	XRPUSDT
flagged	84/84	84/84	83/83	39/39	63/63	84/84

threshold	100	250	500	750
flagged	102/102	126/126	105/105	104/104

Redundancy map (most-flagged pairs across the grid)

pair	cells flagged
`ofi <-> turnover_imbalance`	437
`close <-> vwap`	437
`agg_record_count <-> individual_trade_count`	428
`price_impact <-> volume`	395
`sell_volume <-> volume`	202
`buy_volume <-> volume`	194
`price_impact <-> sell_volume`	92
`buy_volume <-> price_impact`	90

Honest read — in / out

In: the cascade reproduces every known verdict, catches the perfect duplicate in 100% of cells, never false-flags an orthogonal pair, and the verdict is regime-robust under a leakage-free walk-forward.

Out: alignment ≠ value-add. Whether the cascade beats baseline, controls Type-I on unseen pairs, and survives the full Selection gates is Phase 2 (Selection). CODEC/O-info conditional/higher-order nulls and S4 gap-closure are HELD with stated reasons.

read-only ClickHouse · all artifacts under /home/nasimubd/ · S0-frozen params (candidates/codec-foci.md) · raw evidence iterations.jsonl. Deployment 🔒 BLOCKED on Terry.