Iteration 15 — arXiv recovered: three distance cousins complete + the null-machinery backbone, 76 total

Iteration 15 — arXiv recovered 4 HARVESTED · 76 TOTAL NULL-MACHINERY BACKBONE LANDED

2026-07-08 · fifteenth firing · ledger row 20 · harvest batch F13+10-01 · campaign branch loop/matrix-harvest-campaign

In plain words: the archive came back online and paid its debts. The three candidates held back by yesterday's outage all entered with verified papers: the kernel two-sample test (comparing regime distributions without training a classifier — the trap that doomed an earlier method), the regime-switching dependency test that completes yesterday's region-resolved dependence measure, and the optimal-transport two-sample test — which makes three cousins measuring distribution difference with three different geometries (energy, kernel, transport), shipped with a note advising the lab to attack them as a group. The quiet headline is the fourth entry: the stationary bootstrap, a 1994 classic with a maintained implementation. Nearly every instrument in this catalog carries a note saying "needs a block-aware null on our dependent data" — that null now has a citable, implemented backbone instead of an ad-hoc recipe. One padding refusal on the record: the stability suite's flagship paper resists direct resolution, so the suite-level citation stands and no ghost row was added.

The four

id	Instrument	Family	What it adds
H-073	MMD two-sample (hyppo, MIT)	13	kernel geometry for regime homogeneity — no classifier, block-permutation null
H-074	LGC regime-switching test (lg ecosystem)	13	the significance layer on yesterday's region-resolved measure — asymmetric/tail dependency changes between named regimes
H-075	Wasserstein two-sample (POT, MIT)	13	transport geometry — tail mass moving costs more; third cousin, grouped-attack note
H-076	Stationary bootstrap (arch)	10	THE dependence-robust null machinery — the block-null every autocorr note invokes, now citable + implemented

Instrument

Family

What it adds

H-073

MMD two-sample (hyppo, MIT)

kernel geometry for regime homogeneity — no classifier, block-permutation null

H-074

LGC regime-switching test (lg ecosystem)

the significance layer on yesterday's region-resolved measure — asymmetric/tail dependency changes between named regimes

H-075

Wasserstein two-sample (POT, MIT)

transport geometry — tail mass moving costs more; third cousin, grouped-attack note

H-076

Stationary bootstrap (arch)

THE dependence-robust null machinery — the block-null every autocorr note invokes, now citable + implemented

Coverage counters after this firing

Boundary	Minimum	Verified	Screened-pass	Fed	Starvation
STABLE	100	33	23	23	67
FRAGILE↔CONDITIONAL	60	14	0	0	46
CONDITIONAL-scoping	40	3	0	0	37
FAIL	40	13	10	10	27
EXPIRY	40	10	5	0	30
cross-cutting	20	3	0	0	17
Total	300	76	38	33	224

Boundary

Minimum

Verified

Screened-pass

Fed

Starvation

STABLE

100

FRAGILE↔CONDITIONAL

CONDITIONAL-scoping

FAIL

EXPIRY

cross-cutting

Total

300

224

▶ Next iteration

Iteration 16 = FRAGILE↔COND screen S-05. The boundary's unscreened cohort (energy distance, betting two-sample, the stability suite + adjusted measures, RBO, LGC measure + test, MRP, MMD, Wasserstein — recount at screen time, ~10) faces M0/M1; the three-cousins and k_v-bar notes will shape the verdicts. Passers ≥10 would trigger feed batch B-04 (the boundary the sibling M0 table marks as MOST starved: "per-environment invariance counting — YES").

Alternatives: EXPIRY citation-refill (4 pending; unlocks its feed path) · COND-scoping harvest (family 16 tail dependence + EVT leads).

Steering point: no directive = FRAGILE↔COND screen next firing. Directives are ledger-recorded verbatim before taking effect.

Machine-readable record: ledger row 20 + catalog §F13+10-01. WebSearch: blocked this firing (recorded). Pages are append-only.