Iteration 42 — H-021 EILLS EXCLUDED: The Empty-Set Collapse

Iteration 42 — H-021 EILLS EXCLUDED KILL — γ-FRAGILE · EMPTY COLLAPSE ε-FRONTIER WITNESS #3

In plain words: yesterday's promising candidate — the one that folds "predict well" and "stay steady" into a single price — aced its exam but failed the field test. Turned loose on all 130 real features, its answer machine broke in a revealing way: as soon as the steadiness weight is meaningful, it answers "nothing" — for every single feature. The reason is arithmetic, not mystery: on real market data no feature has a perfectly steady relationship with the anchors (their predictive power is tiny — median 2.6%), so the steadiness charge always outweighs the prediction reward and the empty answer is always cheapest. An instrument that gives the same answer for everything measures nothing — the same "dead weight" failure our very first planted control was designed to catch. The silver lining is scientific: this is now the third independent engine to conclude that perfect steadiness does not exist on this market — steadiness is a matter of tolerance, exactly as iteration 22 discovered. The successor is already in the catalog: the tolerance-based relaxation (H-026 lead). Next up: the queue's next candidate, Causal Dantzig.

Preflight (A0): 2026-07-10 01:05 — load1 2.61/32c (≤24) · second claude session visible but light (<2 cores) · 32 GiB avail, si/so=0 · CH + sidecar + kintsugi active → ALL PASS. Sync: main +1 (#601 runbook doc, unrelated); rebased cleanly. One capped run, 0.44 min, 5 workers, readonly=2 loader, watchdog clean, deterministic algebra over real rows only.

Census results (pre-registered rule in eills_m3_census.py)

Check	Reading	Ruling
γ stability (band [1,1000])	Adjacent-γ selection agreement: 1→10 = 0.815 ✗ (bar 0.90) · 10→100 = 1.0 · 100→1000 = 1.0	RULE 1 FIRES — γ-fragile
The collapse behind it	At γ ≥ 10 the selected support is EMPTY for 130/130 subjects: Q(empty)=1 beats every penalized support once γ×penalty exceeds the MSE savings — and the anchors explain only 2.6% (median) of any feature's variance. Selection readout = constant = the C2 dead-weight failure mode. The kill survives any re-draw of the band: fragile at its edge, dead inside it.	DEAD WEIGHT
M3 (moot, recorded)	Only informative readout: poi (price-of-invariance) — worst \|ρ\| = 0.663 vs e-ICP n_acc. Not redundant, but it is a variance-explained gauge, not an invariance instrument.	0.663 < 0.95 — irrelevant given rule 1
Exam-vs-census contrast	The exam (row 52) proved the machinery identifies TRUE invariance where it exists (twin at Q=0 at every γ). The census collapse is therefore a data truth, not a code artifact: no exact invariant anchor relationship exists for ordinary features.	MEASUREMENT ATTACKED & TRUSTED
Verdict (row 53)	EXCLUDE(γ-fragile at census scale — agreement 0.815 < 0.90 at the 1→10 edge; selection constant-empty for γ ≥ 10) → BONEYARD. Successor path on record: ε-relaxed / robust variant with a derived γ (the H-026 invariance-guided-relaxation lead).

Check

Reading

Ruling

γ stability (band [1,1000])

Adjacent-γ selection agreement: 1→10 = 0.815 ✗ (bar 0.90) · 10→100 = 1.0 · 100→1000 = 1.0

RULE 1 FIRES — γ-fragile

The collapse behind it

At γ ≥ 10 the selected support is EMPTY for 130/130 subjects: Q(empty)=1 beats every penalized support once γ×penalty exceeds the MSE savings — and the anchors explain only 2.6% (median) of any feature's variance. Selection readout = constant = the C2 dead-weight failure mode. The kill survives any re-draw of the band: fragile at its edge, dead inside it.

DEAD WEIGHT

M3 (moot, recorded)

Only informative readout: poi (price-of-invariance) — worst |ρ| = 0.663 vs e-ICP n_acc. Not redundant, but it is a variance-explained gauge, not an invariance instrument.

0.663 < 0.95 — irrelevant given rule 1

Exam-vs-census contrast

The exam (row 52) proved the machinery identifies TRUE invariance where it exists (twin at Q=0 at every γ). The census collapse is therefore a data truth, not a code artifact: no exact invariant anchor relationship exists for ordinary features.

MEASUREMENT ATTACKED & TRUSTED

Verdict (row 53)

EXCLUDE(γ-fragile at census scale — agreement 0.815 < 0.90 at the 1→10 edge; selection constant-empty for γ ≥ 10) → BONEYARD. Successor path on record: ε-relaxed / robust variant with a derived γ (the H-026 invariance-guided-relaxation lead).

WHY "NOTHING" IS ALWAYS CHEAPEST (γ ≥ 10, ordinary features): Q(empty) = 1.0 (standardized variance, no penalty) Q(any S ≠ ∅) = (1 − R²_S) + γ · penalty_S (R²_S ≤ ~0.026 median · penalty_S > 0 always, ≥ 0.974 ≥ 10 · penalty because NO exact invariance exists) → Q(S) > 1 the moment γ·penalty > R²_S ⇒ empty wins everywhere ⇒ constant readout ⇒ dead weight THE SAME FACT, THREE ENGINES (ε-frontier, iter 22): e-ICP twins exams → only the byte-identical twin is exactly invariant StabReg (iter 37) → zero surviving stable+predictive teams on real targets EILLS (this iter) → empty-set collapse for all 130 subjects

▶ Next iteration

Next firing: B-01 advances to H-018 Causal Dantzig — entrance exam on the byte-identical forex twin anchor per its harvest CANDIDATE-CATALOG dossier (cheapest-kill-first). Standing lessons applied: oracle-shadow check (row 51), γ/knob dials must be derived or swept in-exam (rows 51–53), and any invariance-exactness assumption meets the ε-frontier head-on. After H-018: H-004 PCMCI+ closes B-01; then B-03 (second STABLE batch). Standing for the operator: the 73 cycle-1 draft statuses (iters 11/13), the Frontier #4 expiry-monitor precondition review, and the ripening STABLE-declaration proposal.

Iteration 42 · 2026-07-10 · EXCLUDE (row 53, BONEYARD +1) — kills are wins · capped (0.44 min) · readonly=2 · zero generated values · append-only · evidence: eills_m3_census.py · eills_m3_census_results.json

Iteration 42 — H-021 EILLS EXCLUDED KILL — γ-FRAGILE · EMPTY COLLAPSE ε-FRONTIER WITNESS #3

Census results (pre-registered rule in eills_m3_census.py)

▶ Next iteration

Census results (pre-registered rule in `eills_m3_census.py`)