Iteration 9 — k-of-N Gate Excluded · Inverted-Signal Discovery: Regime-Stable Coupling Anti-Predicts Orthogonality

Iteration 9 — k-of-N Gate Excluded · Inverted-Signal Discovery GATE EXCLUDED + QUESTION KILLED INVERTED DISCOVERY

What happened in this iteration, in plain words: the counting instrument went for its final certification — and produced the campaign's most interesting result yet. Test one (does adding its score improve the certified panel's forecasts?): no — improvement statistically indistinguishable from zero. Test two (do features it rates "regime-stable" actually keep their orthogonal status?): here came the surprise. The relationship is strong, statistically decisive… and backwards. Features whose behavior is most stable across regimes are the ones most likely to LOSE their orthogonal status. Once you see why, it's obvious: the instrument measures how stably a feature is coupled to the market's core signals — and a stably-coupled feature is stably predictable from those signals, which is the definition of NOT orthogonal. The market's activity dial (volumes, trade counts) is stably coupled — and stably redundant. The engineered features that stay orthogonal do so precisely because their coupling keeps breaking. So: the instrument is honest (its detector demonstrably works), but it answers a different question than the statuses ask. Per the rulebook: gate role rejected, the naive mapping question is formally KILLED, the inverted discovery goes on the books, and the successor asks the semantically exact question — count the per-regime orthogonality verdicts themselves.

Preflight (A0)

2026-07-03 12:19 UTC — load1 3.49 / 32 cores (≤24) · no competing nasimubd jobs · 21 GiB available, si/so≈0 · clickhouse-server active · sidecar active + /health healthy → ALL PASS. One capped run (0.66 min), readonly=2, pre-authorized (03b), zero generated values (03c). Leakage safety: expanding-window k (envs 0..t only per transition — same information timing as the admitted gauges; iteration 8's static k would have leaked future regimes).

The two legs and what they showed

M4 — incremental value (sealed C2'/C3' machinery, apply-once temporal split): baseline (admitted panel) conditional AUC 0.9753 + expanding k_frac conditional AUC 0.9802 ΔAUC +0.0049, 95% CI [−0.0033, +0.0145] → contains 0 → NO admissible lift collinearity vs admitted: max|ρ| = 0.32 (partially overlaps redundancy — consistent with the mechanism below) → as a PANEL GAUGE: EXCLUDE(DEAD-WEIGHT) M5' — real-data power against the persistence label (pre-registered: ρ > 0): ρ(k_frac, stays-orthogonal) on held-out among-orth rows (n=256): ρ = −0.291 95% CI [−0.356, −0.198] pairing-null band [−0.12, +0.12] → the detector WORKS (signal far outside the null band = power demonstrated) → but the direction is INVERTED vs the pre-registered mapping → leg FAILS (apply-once: recorded at the pre-registered criterion, no re-tuning) WHY INVERTED (mechanism, recorded as discovery — LEDGER row 14): k measures stability of the anchors→feature RELATIONSHIP. stably anchored ⇒ stably PREDICTABLE from the core set ⇒ NOT orthogonal. k=9 activity/size family: stably coupled → flips back to redundant. curated bar_* features: orthogonal BECAUSE their coupling keeps breaking (k≤2) → regression invariance and status stability are DIFFERENT AXES. The statuses need per-regime ORTHOGONALITY-VERDICT counting.

Verdicts (frozen grammar)

Object	Verdict	Where
Permutation k-of-N as a panel gauge / status-boundary gate	EXCLUDE(DEAD-WEIGHT) — ΔAUC CI ∋ 0; M5′ pre-registered direction failed	LEDGER row 13 (supersedes row 11's gate-role clause) · BONEYARD
"Does regression-invariance k-of-N ground FRAGILE↔CONDITIONAL?"	KILLED-QUESTION — it measures anchor-coupling stability, which anti-predicts persistence (ρ=−0.291, decisive)	LEDGER row 14
The inverted signal ("stable coupling ⇒ persistent non-orthogonality")	First-class discovery on the books; may re-enter the ladder as a NON-orthogonality predictor candidate	LEDGER row 14 · BONEYARD revisit path (b)
The k measurement + its calibration design (own-permutation null)	Retained report-only; the calibration pattern carries to successors	row 11 measurement clause stands

Object

Verdict

Where

Permutation k-of-N as a panel gauge / status-boundary gate

EXCLUDE(DEAD-WEIGHT) — ΔAUC CI ∋ 0; M5′ pre-registered direction failed

LEDGER row 13 (supersedes row 11's gate-role clause) · BONEYARD

"Does regression-invariance k-of-N ground FRAGILE↔CONDITIONAL?"

KILLED-QUESTION — it measures anchor-coupling stability, which anti-predicts persistence (ρ=−0.291, decisive)

LEDGER row 14

The inverted signal ("stable coupling ⇒ persistent non-orthogonality")

First-class discovery on the books; may re-enter the ladder as a NON-orthogonality predictor candidate

LEDGER row 14 · BONEYARD revisit path (b)

The k measurement + its calibration design (own-permutation null)

Retained report-only; the calibration pattern carries to successors

row 11 measurement clause stands

Technical record

Item	Value
Leakage safety	expanding-window k_frac = k(envs 0..t)/(t+1) per transition epoch t (epochs 0–1 unavailable → cal-median impute); P=500 per expanding run (DERIVED, cost)
Machinery	M4 via the seal-validated `m4_run` (cluster bootstrap B=400, refit pairing-null); M5′ = parameterless Spearman with cluster bootstrap + pairing-null (no thresholds, no new knobs)
Panel	sha-pinned tierA_v3 cells; cal 547 / test 256 among-orth; 100% of test rows carried a k value
Numbers	M4 ΔAUC +0.0049 CI [−0.0033,+0.0145] · max\|ρ\| 0.32 · M5′ ρ −0.291 CI [−0.356,−0.198] vs null ±0.12 · wall 0.66 min
Harness	`k_of_n_admission_legs.py` · `k_of_n_admission_results.json`
Compute / prod impact	one 0.66-min capped run · 10 readonly=2 SELECTs · zero writes outside audit folder + dashboard

Item

Value

Leakage safety

expanding-window k_frac = k(envs 0..t)/(t+1) per transition epoch t (epochs 0–1 unavailable → cal-median impute); P=500 per expanding run (DERIVED, cost)

Machinery

M4 via the seal-validated m4_run (cluster bootstrap B=400, refit pairing-null); M5′ = parameterless Spearman with cluster bootstrap + pairing-null (no thresholds, no new knobs)

Panel

sha-pinned tierA_v3 cells; cal 547 / test 256 among-orth; 100% of test rows carried a k value

Numbers

M4 ΔAUC +0.0049 CI [−0.0033,+0.0145] · max|ρ| 0.32 · M5′ ρ −0.291 CI [−0.356,−0.198] vs null ±0.12 · wall 0.66 min

Harness

k_of_n_admission_legs.py · k_of_n_admission_results.json

Compute / prod impact

one 0.66-min capped run · 10 readonly=2 SELECTs · zero writes outside audit folder + dashboard

▶ Next iteration

Iteration 10 = Frontier #2, refined successor: verdict-stability k-of-N — count the thing the statuses are actually about. In plain words: instead of asking "is the feature's coupling to the core signals stable per regime?" (iteration 9 proved that's the wrong axis — it anti-predicts), ask the literal status question per regime: "was this feature ORTHOGONAL in regime e, yes or no?" — then k = in how many of the 10 regimes the orthogonality verdict held. A feature orthogonal in 9 of 10 regimes is CONDITIONAL material by definition; orthogonal in 2 is FRAGILE. Technically: per-env redundancy verdict (the admitted G0 machinery, worst-cell convention) counted across the 10 environments, with the same own-permutation calibration pattern that survived iteration 8, driven through the ladder (M0–M5′) in one bounded slice — including the M5′ direction test against the persistence label, which for THIS variant has a committed-anchored positive expectation (per-regime orthogonality is literally the label's own ingredient at t). Steering alternatives: full forex grid (fix-plan step 6), forex k-of-N, or F014 (identifiability-sharpened).

Iteration 9 · 2026-07-03 · Frontier #2 admission legs: gate excluded, question killed, inverted discovery recorded · capped read-only · zero synthetic data · append-only