In plain words: yesterday's celebration met today's discipline. The four instruments whose citations were finally resolved โ held at the door fourteen iterations by the no-hearsay rule โ faced the gates at last, and three of the four died there. The adaptive conformal method turned out to duplicate machinery already partially in the lab's stack. The 2007 Bayesian changepoint classic lost to its own robust descendant, which passed earlier precisely because it fixes what breaks on fat-tailed data. The discriminative drift detector's own catalog row had predicted its death (same mechanism as the lab's trusted drift gate) โ the prediction executed. And the drift-adaptive ensemble turned out to be an adaptation system, not a measuring instrument. Only the offline segmenter passed. This looks like a bad day and is actually the system working: the provenance rule delayed these four, and the dedup discipline killed three โ the pipeline's two honesty mechanisms composing exactly as designed. The EXPIRY boundary still lacks its handoff batch, but the refill path is now precise: mechanisms with native statistical guarantees the stack genuinely lacks โ starting with a windowing detector whose false-alarm bounds are printed in its 2007 paper.
The composition of honesty mechanisms
scaffold day: 4 entries, checked code, UNRESOLVED papers
โโโ provenance rule: "no citation, no screening" โ 14 iterations at the door
iteration 18: citations resolved (the debt cleared)
iteration 19: dedup discipline at the gate:
H-011 ACI โ stack already holds the adaptivity โ OUT
H-012 BOCPD โ robust descendant already passed โ OUT
H-013 D3 โ its own row predicted the G5 collision โ OUT
H-014 PWPAE โ adaptation system, not an instrument โ OUT
H-079 PELT โ genuinely new offline slot โ PASS
Neither rule wasted the other's work: unverified entries never reached
the gate; verified duplicates never reached the lab. Kills are wins.
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
11
6 (below threshold)
0
cross-cutting
20
5
0
0
Total
300
79
51
45
โถ Next iteration
Iteration 20 = EXPIRY refill harvest โ native-guarantee mechanisms only. The S-06 kills define the bar: no more stack-adjacent drift detectors. Targets: ADWIN (BifetโGavaldร 2007 โ false-positive/negative bounds printed in the paper, the exact M1 currency; River implementation), Page-Hinkley (1954 classic, resource-efficient in the finance comparison), KSWIN (with its FPR caveat on record), the changepoynt suite mechanisms (SST/density-ratio/FLOSS โ genuinely absent from the stack), and family 18 (mSPRT, sequential ranks). Four solid passers would put EXPIRY at 10 โ B-05.