In plain words: the second delivery went onto the certification lab's intake shelf. Ten vetted redundancy instruments โ the tools that declare "this feature adds nothing new today" โ packaged cheapest-kill-first: instruments whose papers ship their own statistical guarantee go up front (a transfer-entropy variant with an exact formula-based null leads, followed by a 1948 classic with an exact test and a fast modern algorithm), while the two whose trust anchors we had to construct ourselves ride at the back, flagged. The batch also ships an unusual courtesy: a note telling the lab that four of the ten are statistical cousins, so it can attack them as a group and let the first survivor eliminate the rest โ saving compute on our own candidates. With this, the FAIL boundary becomes the second to complete the full harvestโscreenโfeed cycle. The web-search channel answered a crypto-phrased query this firing and contributed five new leads (extreme-value change-points, path-space regime detection, and the standard change-point library) โ parked in the taxonomy for verification.
The pipeline so far โ two boundaries fully cycled
closed-form finite-data null โ no surrogates; no-license flag travels
2
H-031
Hoeffding's D
exact null + O(n log n) CRAN implementation
3
H-032
BergsmaโDassios ฯ*
CRAN active, asymptotic null
4
H-033
Distance correlation
MIT PyPI; block-permutation to wire
5
H-034
dHSIC
CRAN active; joint-independence readout
6
H-055
SKIT (sequential, anytime-valid)
MIT; block-pair sampling design needed
7
H-050
Effective Transfer Entropy
surrogate null is the method โ moderate compute
8
H-035
Copula entropy
M1 borderline โ M2 must confirm null collapse
9
H-041
BROJA-2PID
constructed anchor; discretization setup
10
H-042
O-information
constructed anchor + June-WATCH flag โ last
Group note shipped with the batch: rows 2โ5 are four cousins in adjacent power niches โ grouped M3 attack recommended; first survivor likely kills the rest.
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
10 (B-02)
EXPIRY
40
10
0
0
cross-cutting
20
2
0
0
Total
300
59
20
20
Bookkeeping note (on the ledger): iteration 8 updated the FAIL screened-pass cell but not the Total โ this firing's recompute corrected it. Recompute-not-increment caught the drift, as designed.
โถ Next iteration
Iteration 10 = STABLE harvest resumes (deficit 75 โ still the largest). Options in priority order: family 12 remainder (distribution-shift robustness bounds โ DRO/groupDRO diagnostics beyond the catalogued IRM line), family 20 remainder (IV/anchor-validity diagnostics beyond HSIC-X), or verifying the parked WebSearch leads (local Gaussian correlation, FNETS, sliced Wasserstein). EXPIRY (10 verified) is also approaching a screenable cohort โ its screen batch is the natural iteration-11 slice.
Steering point: no directive = family-12 STABLE harvest next firing. Directives are ledger-recorded verbatim before taking effect.