Iteration 26 — ADMIT #4: e-ICP-fp Crypto Leg · Admitted on Both Substrates

Iteration 26 — ADMIT #4: e-ICP-fp Crypto Leg ADMITTED ON BOTH SUBSTRATES

In plain words: yesterday's certificate covered forex, with a conditional on crypto pending one last test: does the new instrument's evidence trail — how strongly it believes a feature's couplings are regime-stable, using only data up to each moment — actually tell you anything about whether that feature stays orthogonal next regime? The answer came in two parts. Bolted onto the existing three-gate panel it added nothing (the panel is already near ceiling at 0.975 AUC — the same wall the regime counter hit). But on its own, the direction test says yes: features the instrument scores as more invariant ARE more likely to stay orthogonal, with a confidence interval solidly above zero. That is exactly the disjunction the rulebook sealed when the regime counter was admitted the same way. Fourth certificate issued — the STABLE-identification instrument is now admitted on BOTH substrates, and the three crypto features with identified invariant sets graduate from curiosity to actionable STABLE-declaration raw material. One honesty flag: one of the two sub-checks passed by a whisker; the built-in expiry re-tests it automatically next regime.

Preflight (A0): 2026-07-08 13:15 UTC — load1 3.68/32c (≤24) · 23 GiB avail · CH + sidecar active → ALL PASS. Two capped runs (4.5 min each), hermetic crypto pool + committed sha-pinned cells, permutations/bootstraps of real rows only.

Design (leakage-safe)

Per epoch t ≥ 2, the v5 instrument (complete-case pool · standardized design · fold-product J=5/K=99 · degenerate guard) runs over environments 0..t ONLY. Candidate column: log(min_e) over the 15 anchor subsets — the invariance-evidence floor known at time t. Measured on BTCUSDT@250 (the instrument's pool), applied per-feature across combos (the documented iter-9 convention). Panel + machinery = the sealed C2′/C3′ apparatus, panel sha 2ed074c9….

The two legs (sealed row-15/18 disjunction: M4 pass OR M5′b pass)

Leg	Readout	Result
M4 persistence-panel ΔAUC	ΔAUC = −0.0144, CI95 [−0.0456, 0.0] ∋ 0; inside pairing-null band [−0.0239, +0.0025]; baseline already 0.975 AUC	FAIL — no incremental lift over the admitted panel (the same ceiling k_v hit on crypto)
M5′b direction	ρ(log min_e, orth_t1) = +0.1228 on n=256 held-out among-orth rows · combo-bootstrap CI [+0.059, +0.168] > 0 · exceeds pairing-null hi 0.1221	PASS — CI leg decisive; null-exceedance margin razor-thin (+0.0007), recorded honestly
Verdict (row 35)	ADMIT(crypto, 10-window regime grid, role = STABLE invariant-set identification readout, expiry = next completed crypto regime slice) — the row-34 INCLUDE-IF(crypto) is RESOLVED. e-ICP-fp is admitted on BOTH substrates; the STABLE slot is filled stack-wide. The thin null margin is an expiry-review item — the next completed crypto slice re-tests it automatically.

Leg

Readout

Result

M4 persistence-panel ΔAUC

ΔAUC = −0.0144, CI95 [−0.0456, 0.0] ∋ 0; inside pairing-null band [−0.0239, +0.0025]; baseline already 0.975 AUC

FAIL — no incremental lift over the admitted panel (the same ceiling k_v hit on crypto)

M5′b direction

ρ(log min_e, orth_t1) = +0.1228 on n=256 held-out among-orth rows · combo-bootstrap CI [+0.059, +0.168] > 0 · exceeds pairing-null hi 0.1221

PASS — CI leg decisive; null-exceedance margin razor-thin (+0.0007), recorded honestly

Verdict (row 35)

ADMIT(crypto, 10-window regime grid, role = STABLE invariant-set identification readout, expiry = next completed crypto regime slice) — the row-34 INCLUDE-IF(crypto) is RESOLVED. e-ICP-fp is admitted on BOTH substrates; the STABLE slot is filled stack-wide. The thin null margin is an expiry-review item — the next completed crypto slice re-tests it automatically.

Why the disjunction is fair, not lenient: it is the exact pre-existing admission rule under which k_v earned ADMIT #1 (crypto, M4 narrow-miss + M5 decisive) and ADMIT #2 (forex). Changing the rule per-candidate — in either direction — would be the sin. Both the M4 failure and the thin M5′b margin are printed in the ledger row, not smoothed over.

Consequence: the three crypto features with non-empty identified invariant sets (aggression_ratio, bar_ehlers_increment_asymmetry, bar_cox_stuart_trend_z, row 34) are now readouts of an instrument admitted on their own substrate — eligible STABLE-declaration raw material for a future proposal iteration.

▶ Next iteration

Next firing: B-01 continues cheapest-kill-first with H-016 seqICP (sequential/anytime ICP variant) — entrance exam first (the v5 form on the byte-identical forex twin anchor: E1 truth-acceptance + E2 wrong-subset rejection), then the ladder only if it passes. Then H-023 IAS → H-019 IPP → H-025 HSIC-X onward; B-03 after B-01. A STABLE-declaration proposal from the three identified sets is now unlockable and may be taken as a bounded iteration. Standing for the operator: the 73 cycle-1 draft statuses (iters 11/13) and the Frontier #4 expiry-monitor precondition review — its "≥1 promoted feature" precondition now has genuine candidates behind an instrument admitted on both substrates.

Iteration 26 · 2026-07-08 · ADMIT #4 (row 35) — e-ICP-fp admitted on both substrates · capped · zero synthetic data · append-only · evidence: eicp_crypto_m4.py · eicp_crypto_m4_m5b.py · eicp_crypto_m4_m5b_results.json