Iteration 69 — ADMIT #13: the MMD Heterogeneity Meter CERTIFICATE #13 · THIN-MARGIN CAUTION SCOPED · 20 GATE SLOTS

In plain words: the kernel-distance candidate passed its final test — barely, and the record says so out loud. On held-out data it improved forecasts of which features keep their edge, with the entire confidence interval positive — the frozen admission rule is satisfied, so the certificate is issued and the FRAGILE-vs-CONDITIONAL boundary now rests on TWO instruments instead of one. But where every previous certificate beat the "could random re-pairing produce this?" bar by six to eighteen times, this one's lift actually sits inside that bar. Before certifying, the loop attacked its own result the hardest way it knows: splitting the held-out era in two and asking whether the improvement holds in both halves or flips sign (the failure mode that killed a previous candidate). It holds in both — no flip — though most of the lift lives in the earlier half. All of this is written onto the certificate, and the question "should the admission bar itself be raised?" is flagged for the operator rather than answered unilaterally. A bonus discovery: the candidate's direction reading is a near-twin of the calibration meter's — both say features whose behavior shifts most across regimes tend to LOSE their edge, while an older discovery says the most regime-stable features ALSO lose it. Both extremes lose; the middle keeps. That pattern now has five data points.