The Sovereignty Foundation — Research Agreeing on Up

The Correction Algebra, and What Happened When We Tried It — Paper 05 of The Geometry of Meaning
Arc · Forthcoming

Abstract

Four papers of bad news deserve one of repair. We now know the shape of the disease: origins that drift, cones that compress every similarity, frames that were never shared. This paper collects the cures — and they turn out to be humble, mechanical, almost disappointingly ordinary operations, which is exactly what you should want from a repair kit. Slide the origin to the center of the cloud, and half the pathology vanishes: this is mean-centering, the semantic equivalent of moving your zero point from a random field to the center of the Earth. Stretch the squashed directions until the cloud is round: whitening, the same calibration a surveyor applies to a biased instrument. Where the index cannot be touched, standardize each query’s scores in place: the z-score, a pocket-sized frame correction applied at the moment of measurement. None of these is exotic. All of them are the geodesy of the previous papers, translated into linear algebra.

But correction within one space was never the real prize. The series has been walking toward a harder question — the two observers of Paper 01, each holding a private geometry — and here the repair kit produces its most beautiful tool: shared anchors. Two spaces that cannot compare coordinates can still agree to describe everything relative to a handful of common landmarks, exactly as two surveyors who share no map can still both see the same mountain peaks. Represent every point by its relationships to the anchors, and the private axes cancel out of the arithmetic entirely. Agreement stops requiring a shared space. It requires only shared landmarks and honest geometry around them.

And because this series ends in a theory that claims to be empirical, we tried it — in stages, with the bar frozen before each run. Stage one, the mechanism: one space, re-framed — rotated, shifted, noised. Naive coordinate matching fell to chance, 4.4 percent; the same points re-described relative to shared anchors recovered 100 percent, and held 97.8 with rotation and noise. Stage two, the frontier: two genuinely independent models — different builders, different data, different dimensionality, no linear map between them that even type-checks. There the result split, and the split is the finding. The neighborhood structure crossed robustly: nine in ten targets arrived within the top five against a 0.35 percent chance floor — and in the campaign’s boldest test it survived a change of language and model at once, English on one side, Spanish translations embedded by a different organization’s model on the other, median rank 5 of 286 through shared landmarks alone. The exact point crossed about half the time — 56.6 percent against a frozen 60 percent bar, missed twice on held-out data and recorded both times. And the single largest lever in the entire method grid was precisely the correction this paper teaches: mean-centering the profiles roughly doubled cross-model decoding. I will show the runs, the harnesses, and the sealed disproofs alongside the proofs, because a repair kit you cannot watch failing is just another claim.

Two observers never needed the same coordinates. They needed the same landmarks — and the discipline to measure from them. What crosses between two private skies is the shape of the neighborhoods. The exact star, only sometimes — and knowing which is which is the whole discipline.