โ€บNavigation

Iteration 99 โ€” Four-Cousins Quadruple Kill KILLS #39โ€“#42 ยท CERTIFICATE CLASS: 6 MEMBERS ยท SELF-FALSIFICATION #6 CAUGHT

In plain words: four textbook dependence detectors โ€” Hoeffding's D, the ฯ„* sign covariance, distance correlation, and the kernel test dHSIC โ€” were examined together, as the feed recommended. All four SEE the dependence between trading intensity and bar duration loud and clear (4 to 70 times their noise floors). And all four died the same death: the noise floor each one ships with assumes independent samples, and on market bars every one of them cries "signal!" on provably-null data far above its promised 5% โ€” Hoeffding's "exact" null the worst offender on record at 57%, and even distance correlation's block-shuffled null (the repair its own literature suggests) still fails at 24%, because short blocks can't fake long memory. The broken-certificate family the loop first spotted two kills ago now has six members โ€” law-grade. Two bonuses: the exam's own self-check caught a subtle bug in our ฯ„* implementation (algebra says certain patterns score exactly zero; floating-point arithmetic said almost-zero, which the sign function turned into noise) โ€” caught, fixed, re-run, verdict unchanged; and the feed's prediction that these four cousins are interchangeable was refuted โ€” distance correlation turns out to be a genuinely different witness, which makes "distance correlation with the loop's own certified shift-based noise floor" the family's best re-proposal.

Preflight (A0): 2026-07-14 00:30 โ€” load1 4.34/32c ยท no heavy nasimubd jobs ยท 27 GiB avail, si/soโ‰ˆ0 ยท clickhouse active โ†’ ALL PASS. Sync: main +11 โ€” 65-commit rebase conflicted; resolved per the generated-artifact rule as merge 18a0a3c2 (campaign folder OURS ยท other generated nav THEIRS ยท nav rebuilt). PR #599 OPEN. Three capped runs (2.37 + 0.36 + 0.60 min), readonly=2 loader, permutations/shifts/blocks of real values (03c). Operator directive 2026-07-14 recorded (row 112): finish evaluations โ†’ handover (ADR live expiry + flowchart) โ†’ PARK.

Exam results (cousins_exam.py + HV addendum + cousins_exam_v2_taustar.py)

CandidateE2 โ€” own-null validity (THE KILL): pooled FA at nominal 5%, bar 10%E3 reality (own shift null)G0 echoVerdict (row 113)
H-031 Hoeffding's D56.6% โ€” the "exact distribution-free null" is exact only under iid; WORST certificate on record44โ€“70ร—0.977 โ€” G0 DUPLICATE (doubly dead)EXCLUDE ยท KILL #39
H-032 ฯ„* (v2 corrected kernel)11.1% โ€” the closest E2 call on record; the frozen rule fires32โ€“36ร—0.916EXCLUDE ยท KILL #40
H-033 distance correlation24.2% โ€” with the feed-specified block(n1/3)-permutation null: block-length insufficiency โ€” short blocks can't reproduce long-range dependence4.2โ€“4.8ร—0.59 โ€” the family's independent witnessEXCLUDE ยท KILL #41
H-034 dHSIC39.1% โ€” permutation null24โ€“30ร—0.83EXCLUDE ยท KILL #42

Discoveries beyond the kills

#Finding
1ฮต-doctrine self-falsification #6 (caught by the HV gate, fixed in v2): 8 of 24 quadruple orderings of the ฯ„* sign kernel are ALGEBRAICALLY zero but float to ยฑ1e-16 โ†’ sign() emits ยฑ1 noise, inflating ฯ„*(z,z) from the combinatorial 2/3 to 0.846. Tolerance-corrected kernel (|E| โ‰ค 1e-9 โ†’ 0) restores the ceiling exactly. v1 ฯ„* legs superseded append-only.
2Ties discovery: 262/43/215 duplicate duration values per env โ€” the no-ties classical forms (D, ฯ„*) are tie-sensitive; rank-invariance failures recorded as harness caveat, NOT candidate evidence (fed CRAN forms carry tie corrections).
3The feed's mutual-echo prediction REFUTED: the cousins are NOT interchangeable (max sibling echo D~ฯ„* 0.936; dcor 0.50โ€“0.70 vs all) โ€” four different witnesses, one broken certificate class.
4RE-PROPOSAL filed (constructive residue): dcor (and ฯ„*) + the campaign's one CERTIFIED null machinery โ€” the circular-shift null (row 100, FWER 0/49). Real reading, independent witness, valid floor. Queued for the operator, not self-admitted.

The IID-certificate class โ€” six members, LAW-GRADE

#CertificatePromised ฮฑActual FARow
1HAC-t p-values5%43.4%99
2ฯ‡ยฒ(12) closed-form TE null5%16.2%111
3Hoeffding's D "exact" null5%56.6%113
4ฯ„* iid null5%11.1%113
5dcor block(n1/3)-perm null5%24.2%113
6dHSIC perm null5%39.1%113

Proposed standing law (operator ratification pending): independence-assuming nulls โ€” closed-form, permutation, or short-block โ€” are inadmissible p-sources on this substrate. The certified counter-example remains the de-aligned circular-shift null (row 100).

โ–ถ Next iteration

Iteration 99 ยท 2026-07-14 ยท VERDICT ร—4 (row 113) โ€” EXCLUDE ร—4 ยท kills #39โ€“#42 ยท boneyard #54โ€“#57 ยท capped (3.3 min total) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: cousins_exam.py ยท cousins_hv_addendum.py ยท cousins_exam_v2_taustar.py + three results JSONs