The Depth of Almost Nothing: Near-Null Geometry in a Prime–Archimedean System

Two dense N by N matrices are built from the same shared node data. One is arithmetic: it carries the prime powers up to a scale c, each weighted by the von Mangoldt function divided by the square root of q, and placed according to its logarithm. The other is analytic: a smooth archimedean background with no discrete content in it at all. Neither is small, sparse, or nearly degenerate on its own. Their sum is not generic. On an identifiable subspace it is smaller than either summand by dozens to hundreds of orders of magnitude, reaching values below 10^-77 at the configurations studied. That smallness is not a property of the shape of the arithmetic data but of the exact correspondence between weight and position: reassigning which weight sits at which position, with both multisets held exactly fixed, destroys the effect every time it has been tried. This account documents what the phenomenon is and what it is not. The operator is rewritten exactly as a Loewner matrix of divided differences plus an explicit diagonal term belonging to the archimedean part of Weil's explicit formula. Near-nullity then becomes an exact interpolation condition, and the defect that condition measures turns out to be the same constant at every node. An exact Schur reduction compresses the deepest spectral information into a boundary block whose minimal size is measured, across two trajectories with unrelated cutoff laws, to grow with the matrix size rather than with the arithmetic cutoff. The von Mangoldt weight is forced into the open as an explicit summand and turns out to be enormous, cancelling against the smooth channel to roughly forty-six orders of magnitude. The deep eigenvalue is positive at every configuration tested, so the system approaches the boundary of the positive cone from inside. Its depth grows with the matrix size along a law that the association-destroying null does not share, and the growth is shown by direct intervention to be driven by the bulk rather than by the boundary. More than a dozen candidate mechanisms are eliminated against precise, checkable consequences, and several earlier readings of this investigation are retracted in the body rather than quietly dropped. Comparison against elementary kernels unrelated to the construction, including the covariance of Brownian motion, the discrete Brownian bridge, and the Hilbert matrix, places the phenomenon in a regime that is neither generic nor a signature of arithmetic: degeneration disproportionate to the kernel's own scale. No compressed generative mechanism for the deep block is supplied. The account closes on the narrower question the classification leaves open, namely what property of a kernel family decides whether its degeneration is proportionate or disproportionate. All numerical values are regenerated from a single run recording module checksums, library versions, and atom counts, with every eigenvalue confirmed unchanged under doubling of the working precision.