A Prospectively Frozen Media-Field Reading of the 2026 US Midterm Elections: out-of-sample application of the BMPI hidden-reversal protocol, filed before the polls closed

The BMPI protocol asks a narrow question of media data: does the balance of tone between two political blocs move, in the last days of a campaign, in a direction that published polls do not show? This preprint applies the protocol once, prospectively, to a single case — the 2026 US midterm elections — and is filed before any result is known, so that the reading cannot be adjusted afterwards. The rule, the window, the parameters and the external favourite were fixed in a public repository before the final window opened and are reproduced here unchanged. The frozen output is binary and directional: the H4 indicator is ON, with ΔM = +0.0327, meaning that the realised national popular-vote margin is predicted to fall on the Republican side of the polling average. The polls at the moment of filing say D +8.3 to +8.6; they are reported as the benchmark the prediction is scored against, not as an input to it. Every other diagnostic is null. The change-point test returns p = 0.937 and was null on all 45 live daily runs (minimum p = 0.751); the three-day level is −0.035; within the decisive window the ten daily values have a standard deviation of 0.68, so the signal’s own t against zero is 0.15 (p = 0.88); a randomly placed 10-day window in this campaign returns ON 37.9 % of the time; and the verdict reverses to OFF under either single-stream construction of the same data, for reasons traced to the composition of the GDELT streams rather than to any tonal movement. The contribution claimed here is methodological, not predictive. One frozen prediction on one campaign cannot establish that BMPI forecasts elections. What it does establish is a boundary: a normalised media-tone index can produce a positive signal in the absence of any statistically convincing change of regime, and the sign of that signal depends on the composition of the information stream it is computed from. The paper specifies, in advance, the robustness programme and the scoring thresholds for a version 2 to be filed after the certified result.