Iteration 87 — H-092 Diebold–Mariano EXCLUDED KILL #27 · THE NORMALIZATION CANCELS THE SIGNAL · EXAM KILL

In plain words: the classic forecast-comparison test died at the door with the campaign's sharpest mechanistic story. Its per-regime readings were individually excellent — the test feature's predictive contribution registered as clearly significant in every one of the three regimes. But the instrument's job at this boundary is to read the DIFFERENCES between regimes, and there it was worse than blind: its cross-regime spread came in at less than HALF what randomly re-shuffled pseudo-regimes produce. The reason reconciles a puzzle from two iterations ago, when a joint test found regime structure in this exact data at odds of fifteen million to one: the DM statistic divides each regime's average by that regime's own noise level first — and on this data, crisis regimes have proportionally bigger averages AND bigger noise, so the ratio comes out the same everywhere. The division destroys the very signal the instrument claims to measure. The lesson is now a family bar on the boneyard: per-regime declaration instruments must compare raw moments, not self-normalized statistics.