Iteration 17 — Feed batch B-04: twelve instruments into the lab's most-starved boundary, 45 fed

Iteration 17 — Feed batch B-04 published FOURTH HANDOFF · 45 FED INTO THE LAB'S HUNGRIEST SLOT

2026-07-08 · seventeenth firing · ledger row 22 · batch B-04 · campaign branch loop/matrix-harvest-campaign

In plain words: the fourth delivery targets the certification lab's weakest point. The boundary that separates lucky features from reliably-conditional ones is guarded today by exactly one trusted instrument — the lab's own priority table calls it the most starved slot in the machine. This batch offers twelve vetted candidates for it: three distribution-difference tests in three geometries (delivered as a declared group so the lab can let the first survivor eliminate the cousins), a sequential betting test whose evidence stays valid under continuous watching, region-resolved dependence with its significance test, recurrence coupling, the selection-stability pair with its correlated-features fix, rank-trajectory overlap, per-feature factor-loading congruence, and the worst-regime floor. Two entries carry an unusual flag written in the batch itself: they must beat the incumbent — prove value beyond the lab's existing flip-counter — or die at the redundancy gate. Running total: forty-five instruments delivered across four batches; the only screened queue still open is EXPIRY's five, waiting below the handoff threshold for its refill.

Batch scoreboard

B-01 STABLE ×10 (iter 4) invariance testers B-02 FAIL ×10 (iter 9) redundancy detectors B-03 STABLE ×13 (iter 12) universal inference, anchor triad, robustness margins B-04 FRAGILE↔COND ×12 (iter 17) ← the most-starved slot: k-of-N evidence machines ────────────────────────── 45 instruments fed · catalog 76/300 · EXPIRY queue (5) awaits refill

B-04 order (cheapest kill first)

#	id	Instrument	Kill-cost note
1	H-073	MMD two-sample	paper-native null — cheapest
2	H-056	Two-sample by betting	e-value bound; no-license flag
3	H-074	LGC regime-switching test	paper-native bootstrap test
4–5	H-038/075	Energy · Wasserstein	distance-cousins group with row 1 — grouped M3 attack
6	H-071	LGC measure	package-native CIs; pairs with row 3
7	H-049	CRQA	surrogate machinery moderate
8–9	H-068/069	stabm suite + adjusted measures	stability pair; constructed known-answers
10	H-070	Rank-biased overlap	beat-the-incumbent flag: value beyond k_v or die at M3
11	H-044	Tucker's φ per-feature	constructed readout — M2 decides
12	H-072	Minimum Regime Performance	no code (trivial construction); k_v-comparison flag — last

Instrument

Kill-cost note

H-073

MMD two-sample

paper-native null — cheapest

H-056

Two-sample by betting

e-value bound; no-license flag

H-074

LGC regime-switching test

paper-native bootstrap test

4–5

H-038/075

Energy · Wasserstein

distance-cousins group with row 1 — grouped M3 attack

H-071

LGC measure

package-native CIs; pairs with row 3

H-049

CRQA

surrogate machinery moderate

8–9

H-068/069

stabm suite + adjusted measures

stability pair; constructed known-answers

H-070

Rank-biased overlap

beat-the-incumbent flag: value beyond k_v or die at M3

H-044

Tucker's φ per-feature

constructed readout — M2 decides

H-072

Minimum Regime Performance

no code (trivial construction); k_v-comparison flag — last

Coverage counters after this firing

Boundary	Minimum	Verified	Screened-pass	Fed
STABLE	100	33	23	23
FRAGILE↔CONDITIONAL	60	14	12	12 (B-04)
CONDITIONAL-scoping	40	3	0	0
FAIL	40	13	10	10
EXPIRY	40	10	5	0
cross-cutting	20	3	0	0
Total	300	76	50	45

Boundary

Minimum

Verified

Screened-pass

Fed

STABLE

100

FRAGILE↔CONDITIONAL

12 (B-04)

CONDITIONAL-scoping

FAIL

EXPIRY

cross-cutting

Total

300

▶ Next iteration

Iteration 18 = EXPIRY citation-refill + cross-cutting top-up (double-cheap slice). Resolve the four scaffold-era pending citations (ACI, BOCPD, D3, PWPAE — each one Crossref/arXiv call away from screenability) and add the family-10 null-machinery classics (circular block bootstrap, dependent wild bootstrap, HAC/long-run variance — each with maintained implementations). This unlocks the EXPIRY B-05 path AND fills the thinnest absolute boundary (cross-cutting 3/20).

Alternative: COND-scoping harvest — family 16 (tail dependence) is entirely unswept and holds two parked EVT leads.

Steering point: no directive = refill + top-up next firing. Directives are ledger-recorded verbatim before taking effect.

Machine-readable record: ledger row 22 + FRONTIER-FEED.md §B-04 + catalog §B-04 transitions. WebSearch: blocked this firing (recorded). Pages are append-only.