Iteration 43 — H-018 Causal Dantzig Passes the Entrance Exam

Iteration 43 — H-018 Causal Dantzig Passes the Entrance Exam EXAM PASS · ZERO DIALS · M3 NEXT

In plain words: the next candidate hunts steadiness in a place none of our instruments look: the differences between market regimes. Its logic — if a relationship is truly structural, then the way predictor-and-target move together should shift between regimes in exact lockstep with the way the predictors move among themselves; subtract one regime from another and the structural part cancels perfectly. On the ground-truth exam it was emphatic: for the one truly structural relationship (the byte-identical twin), the cancellation is perfect to fifteen decimal places; for every wrong answer the leftover is enormous — a separation of fifteen orders of magnitude, the widest margin any exam has produced. Two extra virtues: it has no tuning dials at all (the trap that consumed the last two candidates cannot even arise), and it is built to tolerate hidden common causes — an angle absent from our entire shelf. Field test next: does it say anything new across all 130 features?

Preflight (A0): 2026-07-10 01:25 — load1 4.4–5.3/32c (≤24) · transient gh proc <1 core · 25 GiB avail; note: ~4.7 GB system swap in use from an earlier non-loop burst (si trickle 20–60 KB/s, so=0) — the capped scope itself runs MemorySwapMax=0 · CH + sidecar + kintsugi active → ALL PASS. Sync: origin/main unchanged. One capped run, 0.52 min, readonly=2 loader, deterministic algebra over real rows only.

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

Leg	Reading	Ruling
E1 identification	Minimizers = exactly the 4 duration-containing supports, defect ≈ 2×10⁻¹⁵ (exact modulo float dust); smallest minimizer = {duration_us} (parsimony rule — the v5 degeneracy lesson)	PASS
E2 rejection	Every duration-free support at defect 2.78–3.35 — ~15 orders of magnitude above the truth tier, the widest exam separation on record; the empty support correctly rejected at 3.35 (cross-regime anchor-covariance shifts are real and large)	PASS
Knob audit (rows 51–53 lesson)	Clean by construction: unregularized closed form, ALL 45 env pairs stacked (no pairing choice), campaign-standard exactness tolerance — nothing to derive, nothing to sweep	ZERO DIALS
Boneyard collision (dossier note)	Distinct failure mode vs the killed Anchor/DRIG named on record: moment-difference algebra, no anchor-duality step — that kill does not transfer. Also does NOT ride the admitted v5 oracle — M3 is genuinely open	NAMED
M2 flag carried forward	Dossier: inference assumes within-env independence → HAC/block correction needed on this substrate — applies to the inference legs (M4/M5′), not to the deterministic census readout	FOR M4/M5′
Verdict (row 54)	EXAM PASS — checkpoint; ladder walk unlocked, no certificate yet

Leg

Reading

Ruling

E1 identification

Minimizers = exactly the 4 duration-containing supports, defect ≈ 2×10⁻¹⁵ (exact modulo float dust); smallest minimizer = {duration_us} (parsimony rule — the v5 degeneracy lesson)

PASS

E2 rejection

Every duration-free support at defect 2.78–3.35 — ~15 orders of magnitude above the truth tier, the widest exam separation on record; the empty support correctly rejected at 3.35 (cross-regime anchor-covariance shifts are real and large)

PASS

Knob audit (rows 51–53 lesson)

Clean by construction: unregularized closed form, ALL 45 env pairs stacked (no pairing choice), campaign-standard exactness tolerance — nothing to derive, nothing to sweep

ZERO DIALS

Boneyard collision (dossier note)

Distinct failure mode vs the killed Anchor/DRIG named on record: moment-difference algebra, no anchor-duality step — that kill does not transfer. Also does NOT ride the admitted v5 oracle — M3 is genuinely open

NAMED

M2 flag carried forward

Dossier: inference assumes within-env independence → HAC/block correction needed on this substrate — applies to the inference legs (M4/M5′), not to the deterministic census readout

FOR M4/M5′

Verdict (row 54)

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

▶ Next iteration

Next firing: H-018 Causal Dantzig ladder — M3 census over the 130-subject Tier-A grid: per-feature moment-difference readouts (minimum achievable defect + identified-support size) vs the admitted stack (kill bar |ρ| ≥ 0.95, worst across readouts). Watch for the two standing failure modes: constant readout (the EILLS dead-weight collapse — will the defect profile vary across features?) and stack-redundancy (the StabReg oracle-shadow — structurally unlikely here: independent machinery). Redundant or degenerate → EXCLUDE, H-004 PCMCI+ closes B-01; clear → M4/M5′ with the HAC/block correction. 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 43 · 2026-07-10 · CHECKPOINT (row 54) — exam PASS · capped (0.52 min) · readonly=2 · zero generated values · zero dials · append-only · evidence: causal_dantzig_exam.py · causal_dantzig_exam_results.json

Iteration 43 — H-018 Causal Dantzig Passes the Entrance Exam EXAM PASS · ZERO DIALS · M3 NEXT

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

▶ Next iteration

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