Iteration 72 — H-074 LGC-Switch EXCLUDED at the Door: a Blind Gauge

Iteration 72 — H-074 LGC-Switch EXCLUDED at the Door KILL #17 · BLIND GAUGE · EXAM KILL

In plain words: this candidate promised something the certified instruments don't have — not just WHETHER a feature's joint behavior differs between regimes, but WHERE (calm center vs violent tails). The exam killed it at its own game: asked to detect the difference between the covid crash and a mid-cycle range on the campaign's canonical pair — a difference the certified kernel meter reads at up to fifty times its noise floor — it could not tell the real regime split from randomly re-shuffled blocks of the same data. Not on any of the three regime pairs. The mechanism is the price of its promise: measuring dependence IN THE TAILS uses only the handful of points that live there, so its noise floor is enormous, and the honest answer drowns. Same failure mode as the campaign's iteration-12 kill: a blind gauge. Two side lessons landed: both of the exam's attempted "bit-exact" certainty claims failed on floating-point technicalities (square-root round-trips and multiplication order) while the one-billionth-of-a-billionth tolerance passed — more proof that exact equality doesn't exist on real hardware, only ε-close. And its sibling candidate (the measure without the test) stays in the queue with this risk pinned to its name.

Preflight (A0): 2026-07-10 16:05 — load1 4.27/32c (falling vs load15 6.30) · external python job <1 core (under bar) · 25 GiB avail, si/so settling ≈0 · all four services active → ALL PASS. Sync: main up-to-date, worktree 0 behind. Two capped runs (0.39 + 0.38 min), readonly=2 loader, block-bootstrap of real values only (03c-compliant).

Exam results — forms v1 + v2, E3 decisive (lgc_switch_exam.py / _v2.py)

Leg	Reading	Ruling
E3 — the boundary role (THE KILL; identical in both runs)	Block-bootstrap regime test on (duration × intensity): covid↔midcycle max-switch 0.119 vs null 0.257 · gold↔covid 0.414 vs 0.499 · gold↔midcycle 0.139 vs 0.645 — zero rejections, where certificate #13 reads the same pairs at 3.5–50× its null	FAIL — blind at its own job
Mechanism	Tail-local correlation uses only the few points living in the tail — the max-switch null is enormous; the WHERE resolution costs more variance than the switch signal carries (F014 blind-gauge mode, iter 12)	STRUCTURAL
E1 v1 → v2 (form-defect history, mine — NOT the kill)	v1 demanded ρ==1.0 bit-exact: fails — √(v·v) does not round-trip (1−2×10⁻¹⁶ observed). v2's replacement invariant (cxy≡vx bit-exact) ALSO fails — w·(x−m)² and (w·(x−m))·(y−m) associate differently. The ε-criterion \|ρ−1\| ≤ 4·eps passed everywhere (max dev 1 eps)	SELF-FALSIFIED ×2 · ε-PASS
E2 — gauge consistency	Twin substituted into the real pair: profiles bit-equal in all envs, both forms (same-path arithmetic IS exact)	PASS
Verdict (row 83)	EXCLUDE(blind gauge — cannot reject its own block null on real regime pairs) — KILL #17, BONEYARD #32

Leg

Reading

Ruling

E3 — the boundary role (THE KILL; identical in both runs)

Block-bootstrap regime test on (duration × intensity): covid↔midcycle max-switch 0.119 vs null 0.257 · gold↔covid 0.414 vs 0.499 · gold↔midcycle 0.139 vs 0.645 — zero rejections, where certificate #13 reads the same pairs at 3.5–50× its null

FAIL — blind at its own job

Mechanism

Tail-local correlation uses only the few points living in the tail — the max-switch null is enormous; the WHERE resolution costs more variance than the switch signal carries (F014 blind-gauge mode, iter 12)

STRUCTURAL

E1 v1 → v2 (form-defect history, mine — NOT the kill)

v1 demanded ρ==1.0 bit-exact: fails — √(v·v) does not round-trip (1−2×10⁻¹⁶ observed). v2's replacement invariant (cxy≡vx bit-exact) ALSO fails — w·(x−m)² and (w·(x−m))·(y−m) associate differently. The ε-criterion |ρ−1| ≤ 4·eps passed everywhere (max dev 1 eps)

SELF-FALSIFIED ×2 · ε-PASS

E2 — gauge consistency

Twin substituted into the real pair: profiles bit-equal in all envs, both forms (same-path arithmetic IS exact)

PASS

Verdict (row 83)

EXCLUDE(blind gauge — cannot reject its own block null on real regime pairs) — KILL #17, BONEYARD #32

ε-frontier witnesses #4–5: exact invariance does not survive even identical-input arithmetic across expression forms — the sqrt round-trip and multiplication associativity each break bit-equality while ε-closeness holds. The campaign's exactness doctrine stands refined: certain cases must be stated as path-identical comparisons (E2-style) or ε-bounded (≤ a few eps), never as cross-expression bit-equality. LGC-pair note: H-071 (the local-dependence MEASURE, no switch test) stays queued with a NAMED tail-variance risk — a distinct construction, so fate transfer does not auto-apply.

▶ Next iteration

Next firing: B-04 advances to H-038 energy distance entrance exam — the second of the three distance cousins (energy/RKHS/transport geometries). The obvious named risk is the H-056 fate: certificate #13's echo (energy distance and MMD are close mathematical relatives — energy distance IS an MMD with a particular kernel). Per the feed's cousins note, the grouped M3 with H-075 Wasserstein is planned: one census run computing both readouts, both judged against the MMD shadow in the same firing — the H-056 kill suggests neither survives, and two cheap kills in one census closes the distance family. Standing for the operator: the M4-bar question (row 80), the H-062 proposal (row 76), the inverted-mechanism investigation (5 corroborations), the 73 cycle-1 drafts, the Frontier #4 review, the STABLE-declaration proposal, and PR #599 — B-04 now 1 admit · 2 kills.

Iteration 72 · 2026-07-10 · VERDICT (row 83) — EXCLUDE · kill #17 · boneyard #32 · capped (0.39 + 0.38 min) · readonly=2 · zero generated values · zero code reuse · append-only · evidence: lgc_switch_exam.py · lgc_switch_exam_v2.py · both result JSONs

Iteration 72 — H-074 LGC-Switch EXCLUDED at the Door KILL #17 · BLIND GAUGE · EXAM KILL

Exam results — forms v1 + v2, E3 decisive (lgc_switch_exam.py / _v2.py)

▶ Next iteration

Exam results — forms v1 + v2, E3 decisive (`lgc_switch_exam.py` / `_v2.py`)