Iteration 41 — H-021 EILLS Passes the Entrance Exam

Iteration 41 — H-021 EILLS Passes the Entrance Exam EXAM PASS · M3 CENSUS NEXT

In plain words: the next candidate answers the invariance question with a completely different engine: instead of testing feature-sets one by one, it folds "predict well" and "stay steady across regimes" into a single price, and whoever minimizes the total price wins. On the ground-truth exam it was flawless — handed a target that IS one of the predictors byte-for-byte, it named exactly that predictor as the answer (total price exactly zero), threw out every wrong answer with a price that GROWS as the steadiness weight is turned up (the penalty amplifies wrongness rather than hiding it), and — the encouraging part — its answer didn't budge as the steadiness weight swept three orders of magnitude, so the dial that doomed other candidates may be harmless here. Also structurally hopeful: unlike the last candidate (killed as a shadow of an existing instrument), this one does NOT ride our certified engine — so the "do you add anything new?" fight is genuinely open. That fight is next.

Preflight (A0): 2026-07-10 00:45 — load1 2.92/32c (≤24) · no heavy nasimubd jobs · 33 GiB avail, si/so=0 · CH + sidecar + kintsugi active → ALL PASS. Sync note: main moved (#600 — operator amendments to the rotation-probe campaign, not this one; contract unchanged); branch rebased cleanly onto main, force-with-lease push (PR pickup verified). One capped run, 0.43 min, readonly=2 loader, deterministic algebra over real rows only — no permutations even needed.

Exam results (pre-registered form, eills_exam.py)

Leg	Reading	Ruling
E1 identification (every γ ∈ {1, 10, 100, 1000})	All Q-minimizers contain `duration_us`; smallest minimizer = exactly {duration_us} at Q = 0.0 (byte-identical twin → exact zero; supersets tie at 0, resolved by the pre-registered parsimony rule — the v5 degeneracy lesson applied cleanly)	PASS
E2 rejection (every γ)	Every duration-free support strictly separated: Q ≈ 0.89–0.96 at γ=0 (pure MSE) → 137–234 at γ=1000 — the penalty amplifies mis-specification. The perfect-fit breakdown that killed causalicp/F014 cannot occur here (no residual-variance division anywhere)	PASS
E3 mechanism-alive (descriptive, not gated)	Invariance penalty at the pooled-OLS solution on a real second target (burstiness) = 0.601 > 0 — the objective has something to trade on real data	ALIVE
γ dial (early read)	Selection invariant across three orders of magnitude on the exam anchor — census-scale behavior still owed at M3 (the row-51 κ-sweep pattern)	PROMISING · VERIFY AT CENSUS
Verdict (row 52)	EXAM PASS — checkpoint; ladder walk unlocked, no certificate yet

Leg

Reading

Ruling

E1 identification (every γ ∈ {1, 10, 100, 1000})

All Q-minimizers contain duration_us; smallest minimizer = exactly {duration_us} at Q = 0.0 (byte-identical twin → exact zero; supersets tie at 0, resolved by the pre-registered parsimony rule — the v5 degeneracy lesson applied cleanly)

PASS

E2 rejection (every γ)

Every duration-free support strictly separated: Q ≈ 0.89–0.96 at γ=0 (pure MSE) → 137–234 at γ=1000 — the penalty amplifies mis-specification. The perfect-fit breakdown that killed causalicp/F014 cannot occur here (no residual-variance division anywhere)

PASS

E3 mechanism-alive (descriptive, not gated)

Invariance penalty at the pooled-OLS solution on a real second target (burstiness) = 0.601 > 0 — the objective has something to trade on real data

ALIVE

γ dial (early read)

Selection invariant across three orders of magnitude on the exam anchor — census-scale behavior still owed at M3 (the row-51 κ-sweep pattern)

PROMISING · VERIFY AT CENSUS

Verdict (row 52)

EXAM PASS — checkpoint; ladder walk unlocked, no certificate yet

Why this candidate is structurally different from the one just killed: H-017's stability screen literally ran the admitted v5 e-value oracle, so its readout was that instrument's shadow (ρ=0.977, row 51). EILLS computes stability as a penalty on per-environment score discrepancy — independent machinery end to end. The M3 redundancy question is therefore genuinely open, not structurally pre-answered.

▶ Next iteration

Next firing: H-021 EILLS ladder — M3 census over the 130-subject Tier-A grid: per-feature selection readout (selected support + per-support Q profile at the swept γ band) vs the admitted stack (kill bar |ρ| ≥ 0.95, worst across readouts), with the γ dial verified at census scale in the same run (the row-51 pattern: the grid is γ-decomposable, one computation supports the whole sweep). Stack-redundant or γ-fragile → EXCLUDE; clear → M4/M5′. 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 41 · 2026-07-10 · CHECKPOINT (row 52) — exam PASS · capped (0.43 min) · readonly=2 · zero generated values · append-only · evidence: eills_exam.py · eills_exam_results.json

Iteration 41 — H-021 EILLS Passes the Entrance Exam EXAM PASS · M3 CENSUS NEXT

Exam results (pre-registered form, eills_exam.py)

▶ Next iteration

Exam results (pre-registered form, `eills_exam.py`)