Iteration 98 โ B-02 OPENS; H-036 Reduced TE EXCLUDED KILL #38 ยท IID-CERTIFICATE CLASS: 2nd MEMBER
In plain words: a new batch opened โ ten instruments aimed at the FAIL boundary, each claiming to catch a kind of redundancy the incumbent rank-correlation gate misses. The first candidate's sales pitch was speed with rigor: a transfer-entropy meter whose null distribution comes from a textbook formula, no expensive resampling needed. The exam attacked the pitch itself, because the loop has been burned by exactly this once before: textbook formulas for "how big is chance noise" are derived assuming independent samples, and market bars are anything but independent. Sure enough: fed provably-null data (real values with their alignment destroyed), the textbook formula cried "signal!" three times too often โ 16% false alarms where it promised 5%. Same disease as the earlier HAC certificate kill, which makes this a two-member failure class worth a standing rule: on this substrate, only noise floors built from the data's own dependence structure (shifts, blocks) can be trusted. The twist: the meter's actual reading is real โ trading intensity genuinely transmits information into the next bar's duration, strongest in the calm regime โ and it is not a duplicate of the incumbent gate. So the mechanism gets a second chance later in the batch under its surrogate-null variant; only the textbook certificate is dead.
Preflight (A0): 2026-07-11 01:05 โ load1 5.90/32c (flat vs 5.96) ยท no heavy nasimubd jobs ยท 28 GiB avail, si/soโ0 ยท clickhouse active โ ALL PASS. Sync: main +0, worktree clean, PR #599 OPEN. One capped run (0.38 min), readonly=2 loader, shifts + block bootstrap of real values (03c).
B-02 opened (row 110) โ FAIL boundary, 10 items, cheapest-kill-first
H-036 Reduced TE โ H-031 Hoeffding's D โ H-032 ฯ* โ H-033 dcor โ H-034 dHSIC (the four-cousins group: grouped M3 economy โ first survivor likely renders the rest DUPLICATE-GAUGE) โ H-055 SKIT โ H-050 Effective TE โ H-035 copula entropy โ H-041 BROJA-2PID โ H-042 O-information. Boundary currency: today-redundancy beyond G0 Spearman; the feed names M3 collinearity as the expected killer.
Exam results (rte_exam.py โ the attack targets the candidate's headline claim)
Leg
Reading
Ruling
E1 โ exactness (counting path)
Self-target TE โก 0.0 exact (conditioning makes the source cell deterministic โ every term nยทlog 1); substitution bit-equal, 3/3 envs
PASS
E2 โ THE ATTACK: closed-form-null validity (THE KILL)
On de-aligned circular-shift TRUE nulls of real values, the ฯยฒโ.โโ (12) bound false-alarms 19/99 covid ยท 12/99 mid ยท 17/99 gold โ pooled 16.2% vs 5% nominal (validity bar 10%). The iid multinomial null is anti-conservative on autocorrelated bars โ the row-99 failure class (HAC-t: 43.4%/5%) claims a second member
FAIL
E3 โ link reality + G0 collinearity (for the record)
Real vs its own shift null: 2.37ร covid ยท 5.96ร midcycle (strongest in the CALM regime โ the inverse of the tail family's profile) ยท 2.50ร gold; G0-echo 0.26 (bar 0.95) โ NOT a duplicate of the incumbent; the boundary's expected killer did not fire
PASS
Verdict (row 111)
EXCLUDE(closed-form iid null invalid โ 16.2% FA at nominal 5%) โ KILL #38, BONEYARD #53 ยท the reading survives: the mechanism re-enters via H-050 Effective TE (surrogate-null, queued row 7) โ no double-tracking ยท M1 license flag recorded (repo has no LICENSE file)
The IID-certificate class โ two sightings (candidate standing law)
#
Certificate
Promised ฮฑ
Actual FA on true nulls
Row
1
HAC-t p-values (family, all-true-nulls)
5%
43.4%
99
2
ฯยฒ(12) closed-form TE null
5%
16.2%
111
Proposed rule for the operator (NOT self-ratified): closed-form nulls derived under independence are inadmissible as p-sources on this substrate; instruments must bring dependence-aware nulls (de-aligned shifts, block resampling) โ the RW shift-based max-T (row 100, FWER 0/49) remains the one certified counter-example.
โถ Next iteration
Next firing: B-02 advances into the four-cousins group โ H-031 Hoeffding's D first (exact distribution-free null + O(n log n): the batch's cheapest classical kill). Exam under the grouped-M3 economy: whichever cousin first survives its entrance exam triggers the group collinearity attack (H-032 ฯ* / H-033 dcor / H-034 dHSIC as siblings โ the sibling rule applies); expected killer remains G0 collinearity, though H-036's echo (0.26) hints the boundary's incumbent is more beatable than the feed assumed. Note: Hoeffding's D's "exact distribution-free null" is ALSO an iid null โ the E2 attack travels. Standing for the operator: the tri-tier ratification (row 105), the 5-sighting tail-invariance pattern (rows 104โ109), the IID-CERTIFICATE class โ now 2 members, proposed standing law (rows 99/111), the RW+MCS pipeline proposal, the H-005 full-window re-proposal, the 03c wild gray zone, the M4-bar question, the H-062 proposal, the inverted-mechanism proposal, the 73 cycle-1 drafts, the Frontier #4 review, the STABLE-declaration proposal, the boneyard counter reconciliation, and PR #599 โ B-02 now 0 admits ยท 1 kill ยท 9 queued.