Iteration 28 — seqICP M3: the Block-Null Correction Is Load-Bearing

Iteration 28 — seqICP M3: Block-Null Correction Load-Bearing M3 CLEAR · CONVERGENT EVIDENCE

In plain words: the label-free invariance test faced the redundancy gate — and the first measurement was a trap the loop caught itself. Run 1 said every feature fails everywhere: a reading so uniform it carries no information at all (a gauge stuck at zero clears "is it correlated with existing gauges?" trivially — and uselessly). The attack on that measurement found the cause: the shuffle-based yardstick ignores that neighbouring bars resemble each other, making ordinary wobble look like regime change. Exactly this hazard was written in the candidate's sourcing dossier months in advance. Run 2 swapped in a neighbourhood-preserving shuffle — and the instrument came alive: five features now read "invariance-rich", the rest reject, and the correlation with every certified instrument is a comfortable 0.25 (kill bar: 0.95). The gate clears — for real this time. Best of all: two of the five invariance-rich features are exactly the ones the newly admitted identifier singled out last week — two instruments with completely different mechanics agreeing on where the stable structure lives. That agreement is the first cross-instrument corroboration the STABLE frontier has ever had.

Preflight (A0): 2026-07-09 — load1 3.33/32c (≤24) · 23 GiB avail · CH + sidecar active · no heavy jobs → ALL PASS. Two capped runs (0.2 min each), hermetic KP pool + committed cells, permutations of real rows only.

Run 1 → attack → Run 2

Step	Result
Run 1 — iid row-permutation null (K=99, shared)	DEGENERATE — n_acc=0 for all 29 features × 15 subsets. Vacuous M3 (constant readout, ρ≡0). NOT reported as a pass.
Attack (verify-before-report)	iid permutation destroys serial dependence → deflated null spread of block means → systematic over-rejection. The harvest dossier's autocorr note predicted precisely this: "block-null calibration at ~2K bars/regime required."
Run 2 — chunk-permutation null (200 chunks × L=103, within-chunk serial structure preserved; §0 doctrine: permutations of real values)	Readout de-degenerates: bimodal n_acc ∈ {0, 15} — 5/29 features all-accept (invariance-rich), rest all-reject. Worst \|ρ\| = 0.252 < 0.95 vs e-ICP-fp min_e/n_acc, k_v, mean redundancy → M3 CLEAR, non-vacuously.

Step

Result

Run 1 — iid row-permutation null (K=99, shared)

DEGENERATE — n_acc=0 for all 29 features × 15 subsets. Vacuous M3 (constant readout, ρ≡0). NOT reported as a pass.

Attack (verify-before-report)

iid permutation destroys serial dependence → deflated null spread of block means → systematic over-rejection. The harvest dossier's autocorr note predicted precisely this: "block-null calibration at ~2K bars/regime required."

Run 2 — chunk-permutation null (200 chunks × L=103, within-chunk serial structure preserved; §0 doctrine: permutations of real values)

Readout de-degenerates: bimodal n_acc ∈ {0, 15} — 5/29 features all-accept (invariance-rich), rest all-reject. Worst |ρ| = 0.252 < 0.95 vs e-ICP-fp min_e/n_acc, k_v, mean redundancy → M3 CLEAR, non-vacuously.

Convergent evidence (report-only)

The five invariance-rich features: price_impact, bar_categorical_recurrence_rate, bar_sign_markov_flux, bar_ehlers_increment_asymmetry, bar_cox_stuart_trend_z. The last two are 2 of the 3 features e-ICP-fp produced identified invariant sets for (row 34) — agreement across mechanically unrelated instruments (env-labelled e-value fold-products vs label-free time-block statistics). Disagreement on aggression_ratio (e-ICP identifies; seqICP rejects) recorded. Standing rule for this candidate: crypto runs REQUIRE the block-aware null.

▶ Next iteration

Next firing: seqICP ladder continues — M4/M5′b crypto legs with the (leakage-safe, expanding) binary invariance flag as candidate column via the sealed machinery; the bimodal readout makes a crypto persistence kill plausible — a clean scoped EXCLUDE if so. Then the forex head-to-head vs e-ICP-fp decides the candidate's fate (its survival claim: env-free operation). Independently unlockable: the first STABLE-declaration proposal — now with cross-instrument corroboration on two features. Standing for the operator: the 73 cycle-1 draft statuses (iters 11/13) and the Frontier #4 expiry-monitor precondition review.

Iteration 28 · 2026-07-09 · CHECKPOINT (row 37) — M3 CLEAR (block-null form) · capped · zero synthetic data · append-only · evidence: seqicp_ladder_m3.py · seqicp_ladder_m3_blocknull.py · result JSONs