Iteration 35 — H-025 HSIC-X Passes the Entrance Exam

Iteration 35 — H-025 HSIC-X Passes the Entrance Exam EXAM PASS · ALWAYS-ALARM RISK

In plain words: a different kind of candidate: not another stability meter, but a guard for the guards. Every invariance instrument in the stack leans on three anchor signals — and nothing checks whether the anchors themselves are trustworthy. This candidate tests exactly that: is an anchor secretly entangled with what the model failed to explain? Its exam had two parts. With the right model (zero error), it correctly raised no alarms. With a deliberately broken model (the true driver omitted), it correctly caught the omitted driver hiding in the errors. Pass — but the loop wrote down the worry before moving on: in the broken-model test it flagged everything, not just the planted culprit. Plausibly real (it sees nonlinear entanglement ordinary regression can't remove), but a smoke detector that goes off for every meal is worse than none. Next firing runs the census: across 130 real features, how often does this alarm ring? Ring-always = dead weight, exactly how the campaign killed a candidate in its first week.

Preflight (A0): 2026-07-12 — load1 4.10/32c (≤24) · 17 GiB avail · CH + sidecar active · no heavy jobs → ALL PASS. One capped run (4.47 min), complete-case real rows, chunk permutations of real values (serial-dependence-aware, per the dossier's block-HSIC note).

Exam design (pre-registered, guard-role form)

Leg	Design	Result
E1 no false alarm	TRUE model (target ~ duration_us, residuals exactly 0 → degenerate guard)	✓ zero flags on all three anchors
E2 power	OMITTED-VARIABLE model (target ~ intensity only) — the omitted true parent must be flagged	✓ duration_us flagged at p=0.01 (HSIC 0.037, strongest of the three)
Engine	Gaussian-kernel HSIC, median-heuristic bandwidths, stride subsample n=2000, K=99 chunk permutations (blocks of 20)	runs clean under cap
Verdict (row 44)	EXAM PASS — checkpoint; ladder next

Leg

Design

Result

E1 no false alarm

TRUE model (target ~ duration_us, residuals exactly 0 → degenerate guard)

✓ zero flags on all three anchors

E2 power

OMITTED-VARIABLE model (target ~ intensity only) — the omitted true parent must be flagged

✓ duration_us flagged at p=0.01 (HSIC 0.037, strongest of the three)

Engine

Gaussian-kernel HSIC, median-heuristic bandwidths, stride subsample n=2000, K=99 chunk permutations (blocks of 20)

runs clean under cap

Verdict (row 44)

EXAM PASS — checkpoint; ladder next

The worry, in red ink: in E2 all three anchors were flagged at p=0.01 — plausibly genuine nonlinear/heteroskedastic entanglement (OLS removes only the linear part), but there is no ground truth for the other two under this leg, so specificity is untested. An always-alarm guard is dead weight — the same C2 failure mode that killed bootci_w in week one. The ladder's first leg is therefore a flag-rate census: all-anchor models across the 130 subjects; ~100% flag rate = EXCLUDE(dead weight), graded = informative guard.

▶ Next iteration

Next firing: HSIC-X flag-rate census + M3 — per-feature anchor-validity readings (target ~ all 3 anchors, per-anchor HSIC flags) across the 130-subject grid; the flag-rate distribution decides dead-weight vs informative; if informative, correlation attack vs the five-instrument stack. Then B-01 continues with H-017 StabReg. 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 (four certified forex instruments + ∅-invariance lens + shape-shift warning list).

Iteration 35 · 2026-07-12 · CHECKPOINT (row 44) — exam PASS, worry recorded · capped · zero synthetic data · append-only · evidence: hsicx_exam.py · hsicx_exam_results.json