What Do Multi-View Photometric Residuals Actually Measure?

A shiny surface looks different from different viewpoints. A matte one does not. Several recent Gaussian-splatting methods turn that observation into supervision: they measure how much a surface changes across views and use it as a prior on reflection strength. The signal is free, since it is already sitting in a reconstruction you have. We asked what it actually measures. On 64 hand-annotated regions across nine views of four indoor captures, it is mostly albedo, then geometric error, with material a distant third. It separates shiny from matte at $1.6\times$, and its correlation with material class cannot be told apart from zero. Among matte regions alone, where there is nothing about material to find, its log variant correlates $+0.71$ with image texture and $-0.66$ with brightness, while the linear variant has intervals that span zero at $n{=}20$. The trouble with checking this on photographs is that nobody knows the true $\rho_s/\rho_d$ of a kitchen counter. So we built rooms where we do: ray traced, every material chosen by us, written out as COLMAP captures and reconstructed by an ordinary pipeline. Handed \emph{perfect} geometry, the albedo-normalised residual correlates $-0.16$ with true material once albedo is held fixed, and in a second room built so albedo and $\rho_s/\rho_d$ are uncoupled it correlates $+0.02$ with no control at all. The un-normalised residual keeps $+0.63$ in both, so the correction real photographs force on us is the thing that destroys the signal, and normalising harder does not help, because the nuisances are partly caused by the material we are trying to measure. One quantity does track the truth well, the view-dependent energy already in the fitted spherical harmonics ($+0.89$ with albedo held fixed), but it does not transfer to real captures. A second published cue, SSR-GS's Reflection Score, is that un-normalised residual, so our controls would destroy what it uses. Six attempts of our own to rescue a larger effect are also reported. None worked.