Iteration 8 — FAIL screen batch S-02: 10 pass, 1 duplicate out, feed threshold triggered

Iteration 8 — FAIL screen batch S-02 10 PASS · 1 OUT FEED THRESHOLD TRIGGERED (B-02)

2026-07-07 · eighth firing · ledger row 13 · screening batch S-02 · campaign branch loop/matrix-harvest-campaign

In plain words: interrogation round two — this time for the FAIL boundary, the instruments that declare a feature redundant today. A subtlety made this screen harsher than the first: the certification lab already trusts a redundancy gauge (the Spearman gate), so "measures dependence" is not enough — each candidate had to name dependence the incumbent is blind to. Ten did: non-monotone relationships, dependence invisible to rank statistics, whole-group redundancy no pairwise check can see, information that decomposes into unique-vs-shared parts, time-lagged directional flow, and dependence that only emerges as evidence accumulates. One was shown the door before wasting the lab's time: a conditional-independence test whose slot is already occupied by a trusted gate — the exact spot where a boneyarded predecessor died making the same argument. The passers also ship with a candid gift to the lab: four of them are cousins occupying adjacent niches, so the note says "attack them as a group — whichever survives first probably kills the rest." Second handoff batch is due next firing.

The screen (M0/M1 verbatim, default = OUT, incumbent-aware)

11 screenable FAIL entries │ ▼ M0: name the failure mode — SPECIFICALLY one the incumbent G0 Spearman gate misses 1 OUT (DUPLICATE-ON-PAPER): H-039 GCM — conditional-redundancy slot already held by trusted gate G2; boneyarded RCIT died there with the same class of argument. Hardness theorem kept as citation anchor. │ ▼ M1: matched null / error bound / known-answer — constructible on OUR data 0 OUT — but three passers carry borderline flags (constructed anchors, estimator bias) │ ▼ 10 SCREENED-M0M1-PASS ──► FAIL feed threshold (≥10) TRIGGERED → batch B-02 next firing

The ten passers

id	Instrument	The blindspot it covers	M1 anchor
H-031	Hoeffding's D	non-monotone dependence	exact distribution-free null (paper-native)
H-032	Bergsma–Dassios τ*	dependence invisible to rank correlations	asymptotic null + block permutation
H-033	Distance correlation	nonlinear dependence	block-permutation null
H-034	dHSIC	joint d-variable dependence — set-level	permutation/gamma null (paper-native)
H-035	Copula entropy	MI-magnitude redundancy, marginal-free	bias-matched shuffle null (borderline — M2 decides)
H-036	Reduced Transfer Entropy	time-lagged directional dependence	closed-form finite-data null — strongest in cohort
H-041	BROJA-2PID	WHICH share of information is redundant vs unique	duplicate-feature known-answer (constructed — M2 decides)
H-042	O-information	group-level synergy-vs-redundancy sign	constructed anchor + June-WATCH context flag
H-050	Effective Transfer Entropy	directional flow at small samples, bias-corrected	shuffle-surrogate null IS the method
H-055	SKIT	dependence that emerges over time	e-value/Ville bound (paper-native)

Instrument

The blindspot it covers

M1 anchor

H-031

Hoeffding's D

non-monotone dependence

exact distribution-free null (paper-native)

H-032

Bergsma–Dassios τ*

dependence invisible to rank correlations

asymptotic null + block permutation

H-033

Distance correlation

nonlinear dependence

block-permutation null

H-034

dHSIC

joint d-variable dependence — set-level

permutation/gamma null (paper-native)

H-035

Copula entropy

MI-magnitude redundancy, marginal-free

bias-matched shuffle null (borderline — M2 decides)

H-036

Reduced Transfer Entropy

time-lagged directional dependence

closed-form finite-data null — strongest in cohort

H-041

BROJA-2PID

WHICH share of information is redundant vs unique

duplicate-feature known-answer (constructed — M2 decides)

H-042

O-information

group-level synergy-vs-redundancy sign

constructed anchor + June-WATCH context flag

H-050

Effective Transfer Entropy

directional flow at small samples, bias-corrected

shuffle-surrogate null IS the method

H-055

SKIT

dependence that emerges over time

e-value/Ville bound (paper-native)

Coverage counters after this firing

Boundary	Minimum	Verified	Screened-pass	Fed
STABLE	100	25	10	10 (B-01)
FRAGILE↔CONDITIONAL	60	6	0	0
CONDITIONAL-scoping	40	3	0	0
FAIL	40	13	10	0
EXPIRY	40	10	0	0
cross-cutting	20	2	0	0
Total	300	59	20	10

Boundary

Minimum

Verified

Screened-pass

Fed

STABLE

100

10 (B-01)

FRAGILE↔CONDITIONAL

CONDITIONAL-scoping

FAIL

EXPIRY

cross-cutting

Total

300

▶ Next iteration

Iteration 9 = publish feed batch B-02 (FAIL boundary) — the rule fires it: FAIL reached 10 screened passers. Cheapest-kill-first ordering will put the paper-native-null tests up front (Reduced TE's closed-form null, Hoeffding's exact null, SKIT's e-value bound), the constructed-anchor and borderline entries mid-pack, and the four-cousin group note travels with the batch so the lab can run its collinearity attack efficiently.

Steering point: no directive = B-02 publishes next firing. Directives (reorder, hold, resume harvesting instead) are ledger-recorded verbatim before taking effect.

Machine-readable record: ledger row 13 + catalog §S-02. WebSearch: blocked this firing (finance-phrased query; pattern recorded). Pages are append-only.