Iteration 36 — ADMIT #8: HSIC-X, the Anchor-Validity Guard

Iteration 36 — ADMIT #8: HSIC-X, the Anchor-Validity Guard ADMIT (FOREX) · GRADED GUARD

In plain words: the smoke-detector question got its answer: it does not ring for every meal. Run across all 130 features, the alarm produced a genuinely graded picture — 37 features fully clean (their stability evidence rests on untainted anchor signals), 19 partially entangled, and 74 where every anchor is tangled up with what the models can't explain. That last group's invariance certificates now carry a standing caveat; the clean 37 are the bedrock. Redundancy check: 0.44 at worst — the most independent reading of any certified instrument, which makes sense: it watches the evidence, not the features. Eighth certificate issued. The instrument shelf now has a proper hierarchy: four stability meters on top, and underneath them, for the first time, a guard that checks whether the ground they stand on is solid.

Preflight (A0): 2026-07-13 — load1 4.25/32c (≤24) · 17 GiB avail · CH + sidecar active · no heavy jobs → ALL PASS. One capped run (2.13 min; trace-identity engine, NSUB=1000 census scale, K=99 shared chunk-null), committed forex grid + sha-pinned cells, permutations of real rows only. Decision rule pre-registered in-script.

The census (130 subjects, target ~ all 3 anchors, per-anchor flags)

Reading	Value	Ruling
Flag distribution	0 flags → 37 · 1 → 4 · 2 → 15 · 3 → 74 (all-flag rate 0.569)	GRADED — far under the 0.95 dead-weight bar; the row-44 worry resolves in the candidate's favor
M3 worst correlation vs the admitted stack	0.442 (min_p vs seqICP) · ≈0 vs k_v (0.07) and redundancy (−0.085)	CLEAR — the most orthogonal readout of any admitted candidate
What it means	37 features: anchor set validity-clean → their invariance certificates rest on solid ground. 74 features: all anchors residual-entangled → standing caveat on their invariance evidence. A guard on the evidence, not another stability meter.
Verdict (row 45)	ADMIT(forex, 10-env EURUSD@5 grid, role = anchor-validity guard — flags features whose invariance evidence rests on residual-entangled anchors, expiry = next completed forex regime slice) — ADMIT #8. Feed row H-025 consumed.

Reading

Value

Ruling

Flag distribution

0 flags → 37 · 1 → 4 · 2 → 15 · 3 → 74 (all-flag rate 0.569)

GRADED — far under the 0.95 dead-weight bar; the row-44 worry resolves in the candidate's favor

M3 worst correlation vs the admitted stack

0.442 (min_p vs seqICP) · ≈0 vs k_v (0.07) and redundancy (−0.085)

CLEAR — the most orthogonal readout of any admitted candidate

What it means

37 features: anchor set validity-clean → their invariance certificates rest on solid ground. 74 features: all anchors residual-entangled → standing caveat on their invariance evidence. A guard on the evidence, not another stability meter.

Verdict (row 45)

ADMIT(forex, 10-env EURUSD@5 grid, role = anchor-validity guard — flags features whose invariance evidence rests on residual-entangled anchors, expiry = next completed forex regime slice) — ADMIT #8. Feed row H-025 consumed.

▶ Next iteration

Next firing: B-01 advances to H-017 Stabilized Regression — entrance exam first (v5 form); dossier claim: feature-SET-level stability+predictiveness selection (M1-borderline flag travels: its power anchor is constructed, not paper-native — extra scrutiny at the exam). Remaining queue after it: H-021 EILLS → H-018 Causal Dantzig → H-004 PCMCI+; then B-03. 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 — now further sharpened: candidate declarations can require a CLEAN anchor-validity reading (the 37-feature bedrock) before any STABLE stamp.

Iteration 36 · 2026-07-13 · ADMIT #8 (row 45) · capped · zero synthetic data · append-only · evidence: hsicx_census_m3.py · hsicx_census_m3_results.json