In plain words: a perfect score that proves a process point. Two firings ago, the EXPIRY screen went one-for-five โ stack-duplicates and frameworks died at the gate. That massacre wrote a precise shopping list ("mechanisms with printed guarantees or genuinely absent machinery only"), the refill shopped against it, and today all five refill entries passed: the adaptive-windowing detector with its theorem-grade false-alarm bounds, the 1954 inspection classic, and three mechanisms the lab's stack simply doesn't have (density-ratio, spectral-subspace, matrix-profile). Three carry honest borderline notes โ their guarantees are about estimation, not testing, so hand-built permutation checks travel with them for the lab's real-data gate. The contrast โ 1/5 then, 5/5 now โ is the campaign's cheapest lesson on the record: harvest against the screen's own bar and the screen stops being a slaughterhouse. EXPIRY now has eleven vetted passers; the fifth handoff batch fires next, after which every substantive boundary except regime-scoping will have delivered to the lab.
Then vs now
S-06 (iter 19): scaffold-era cohort, harvested BEFORE the bar existed
โ 1 PASS / 4 OUT (3 stack-duplicates, 1 framework)
S-07 (iter 21): refill cohort, harvested AGAINST the S-06 bar
โ 5 PASS / 0 OUT
Lesson (recorded): the screen's kills are the next harvest's specification.
The five passers
id
Instrument
M1 anchor
H-080
ADWIN (river)
printed FP/FN bounds โ strongest in cohort
H-081
PageโHinkley (river)
ARL theory; calibration-baseline role
H-082
RuLSIF (changepoynt)
constructed block-permutation null โ M2 decides
H-083
SST (changepoynt)
constructed Hankel-shuffle null + planted change
H-084
FLOSS (stumpy)
near-native: the paper's no-change arc-curve is the reference
Coverage counters after this firing
Boundary
Minimum
Verified
Screened-pass
Fed
STABLE
100
33
23
23
FRAGILE↔CONDITIONAL
60
14
12
12
CONDITIONAL-scoping
40
3
0
0
FAIL
40
13
10
10
EXPIRY
40
16
11 (pending)
0
cross-cutting
20
5
0
0
Total
300
84
56
45
โถ Next iteration
Iteration 22 = publish feed batch B-05 (EXPIRY, 11 entries) โ the fifth boundary handoff. Cheapest-kill-first: the printed-bounds detectors up front (ADWIN, E-detectors, the confidence-sequence family), the constructed-anchor mechanisms behind, the Frontier-#4-flagged conformal martingale with its coordination note. After B-05, every substantive boundary except CONDITIONAL-scoping (3/40, family 16 unswept) will have fed the lab โ making COND-scoping's first dedicated harvest the clear iteration-23 slice, alongside the flagged families-sweep audit.
Steering point: no directive = B-05 publishes next firing. Directives are ledger-recorded verbatim before taking effect.
Machine-readable record: ledger row 26 + catalog ยงS-07. WebSearch: blocked this firing (recorded). Pages are append-only.