\section{The idea: noise next to the light cone, read by an atom} \label{sec:idea} A classical field that carries a force between quantum bodies must be noisy, and the noise is at least as large as the response it carries. Where the response of gravity is known, the noise is therefore bounded below. Lorentz invariance, or the propagator of gravitational waves, places a known response next to the light cone at X-ray frequencies, and an atom turns noise there into photons that a detector counts. \emph{The trade-off.} A classical gravitational field that responds to quantum matter keeps a record of where the matter is. Positivity of the dynamics then demands two things at once: the record diffuses, and the matter decoheres, and the two cannot both be small. In one form, valid for any diffusion, the total noise that matter feels on a given source is at least the response of the mean field on that source (Eq.~\eqref{eq:response-form}). At zero frequency this is the known statement that the least noise of a Newtonian classical field is the Di\'{o}si--Penrose kernel~\cite{TilloyDiosi2017}. We measure every noise in units of that kernel and call it $s$: the static Newtonian minimum is $s=1/k^2$ at wavenumber $k$. \emph{The picture.} Figure~\ref{fig:strip} shows the plane of frequency $\omega$ and wavenumber $k$ of the noise. Laboratory tests of Newton's law are static: they sit on the axis $\omega=0$, at $k$ between $(1~\mathrm{m})^{-1}$ and $(1~\mathrm{mm})^{-1}$. A Lorentz-invariant kernel depends on $\omega$ and $k$ only through $u=k^2-\omega^2$, so it is constant on each hyperbola of fixed $u$, and each hyperbola runs from the static axis up along the light cone. An atom that emits a photon of energy $E$ draws on the noise at the frequency $\omega=E/\hbar$. On that line the tested band, $u$ from $1$ to $10^{6}~\mathrm{m^{-2}}$, is a strip just outside the cone, at wavenumbers that exceed $\omega/c$ by the relative amount $u/2k^2$ to leading order. Noise on the strip has the wavelength of the photon, not of the laboratory. What Newton's law fixes at a millimetre, Lorentz invariance repeats at every X-ray energy. \begin{figure}[t] \centering \begin{tikzpicture}[>=stealth,scale=1.25] % tested band between two hyperbolae k^2 - w^2 = u \fill[gray!30] plot[domain=0:2.3124,variable=\t,samples=60] ({0.8*cosh(\t)},{0.8*sinh(\t)}) -- plot[domain=1.4926:0,variable=\t,samples=60] ({1.9*cosh(\t)},{1.9*sinh(\t)}) -- cycle; \draw[thin] plot[domain=0:2.3124,variable=\t,samples=60] ({0.8*cosh(\t)},{0.8*sinh(\t)}); \draw[thin] plot[domain=0:1.4926,variable=\t,samples=60] ({1.9*cosh(\t)},{1.9*sinh(\t)}); % axes \draw[->] (0,0) -- (6.2,0) node[below] {$k$}; \draw[->] (0,0) -- (0,4.5) node[left] {$\omega$}; % light cone \draw[thick] (0,0) -- (4.0,4.0) node[above] {light cone, $u=0$}; % static tests \draw[line width=2.2pt] (0.8,0) -- (1.9,0); \node[below,align=center,font=\small] at (1.35,-0.08) {static tests of\\Newton's law}; % X-ray line \draw[dashed] (0,3.2) -- (6.0,3.2); \node[left,font=\small] at (0,3.2) {$E/\hbar$}; \draw[line width=2.2pt] (3.298,3.2) -- (3.722,3.2); \draw[<-] (3.56,3.12) -- (4.45,2.55) node[below right,align=left,font=\small,inner sep=1pt] {the strip:\\same $u$, photon frequency}; % boost arrow, outside the band \draw[->,thick] (2.45,0.95) to[bend right=8] (3.25,2.15); \node[right,font=\small] at (2.95,1.45) {boost}; \node[font=\small] at (4.9,0.55) {$u>0$ (spacelike)}; \end{tikzpicture} \caption{The frequency--wavenumber plane of the noise (schematic, not to scale). Static tests of Newton's law fix the response on the thick segment of the axis $\omega=0$. A Lorentz-invariant kernel is constant on each hyperbola of fixed $u=k^2-\omega^2$, so the tested band (grey) continues along the light cone. An atom that emits a photon of energy $E$ draws on the noise at frequency $E/\hbar$ (dashed), where the band is a thin strip just outside the cone. The tensor argument does not use the grey band: the response of general relativity has its own pole on the cone.} \label{fig:strip} \end{figure} \emph{First argument: the scalar strip.} The direct route fails, and the reason is instructive. In general relativity the Newtonian potential obeys a constraint, not a wave equation: its mean response to a conserved source has no pole on the light cone (Eq.~\eqref{eq:shield}), so ``noise at least as large as the response'' gives slow matter nothing there. The covariant theory supplies the noise by another route. Its diffusion weight, with strength $\alpha$, makes the conformal part of the metric a dynamical random field whose variance on the strip grows as $1/(\alpha u^2)$. Positivity ties the direct decoherence of matter to the same weight, in proportion to $\alpha$. Matter is therefore caught between two channels, one that grows and one that falls with $\alpha$, and the sum is at least twice their geometric mean, in which $\alpha$ cancels. The result is a floor, $s\geq\tfrac23|f|/u$ (Theorem~\ref{thm:floor}), with $|f|=1$ wherever Newton's law holds. It is two thirds of what a classical field coupled to the trace alone gives by the direct route. Integrated over the tested band it is $\tfrac23\ln10^{6}=9.2$, the same at every X-ray energy. An atom reads this noise because its matter and its field respond differently. Next to the cone the fluctuating metric is conformal. A slow particle feels a conformal mode through its rest mass and is accelerated; Maxwell's equations are conformally invariant and do not see it. The charge is shaken relative to its own field and radiates. A neutral atom shaken as a whole radiates through the difference between its nuclear charge and its electronic form factor, which cancels only at long wavelength; from 10 to 140~keV the atomic factor lies between a quarter and one. Our likelihood fit to the spectrum of XENONnT's search for such photons bounds the integrated noise at $7.7\times10^{-8}$ in the same units, a factor of at least $1.1\times10^{8}$ below the floor (median expected $6.6\times10^{7}$). Within the Gaussian class at linear order, three things defeat this argument. The floor needs a diffusion weight that is built from Lorentz-invariant projectors; a preferred frame in the scalar weight breaks that structure, and the floor is then not shown: the one completion of this kind that the theory's authors work out survives the scalar strip (Section~\ref{sec:cp}). The transfer from the static axis to the strip needs Lorentz invariance of the scalar sector up to boosts of about $10^{7}$. A scalar mass above $(0.34~\mu\mathrm{m})^{-1}$, which would have to be non-perturbative, lowers the floor below the limit. Regulators and running couplings that depend on $p^2$ alone defeat nothing, because they are functions of $u$ and the floor holds at each $u$. \emph{Second argument: the tensor strip.} The pole that the Newtonian potential lacks is not absent from general relativity; it sits in the transverse-traceless part of the response, where gravitational waves propagate. On a transverse-traceless source next to the cone the response is $2/u$, and here the direct route works: the noise is at least the response, $s\geq2/u$ per polarisation, for every positive diffusion weight, with or without a frame (Theorem~\ref{thm:tensor}). Nothing is carried over from static tests. The input is that the mean tensor field of the theory is that of general relativity at keV momenta, which is an assumption: gravitational-wave observations test it only at far lower frequencies. That an atom at rest should source a tensor mode is not obvious, since the stress of its moving particles is of order $v^2/c^2$. The stress of its Coulomb field is not small. The field energy is part of the atom's mass, the Maxwell stress is of the same order, and it is part of the conserved stress tensor, to which we assume the classical field couples. Tensor noise acting on the Coulomb field produces photons: at its floor, at no less than the rate of trace noise of strength $\XTT/u$, with a conversion ratio $\XTT$ that rises from $1.6$ at 5~keV to $9.6$ at 140~keV (Proposition~\ref{prop:tt-emission}). The resulting margin is $1.7\times10^{9}$ (median expected $9.8\times10^{8}$). It is an estimate, not a bound: the conversion is computed for an isolated atom with its elastic structure factor. Two things defeat the tensor argument and leave the scalar argument standing. One is a tensor sector whose pole has left the light cone at keV momenta, which for a Lifshitz dispersion with $z=3$ requires a scale below about 10~eV. The other is a matter coupling that does not see the stress of the atom at all. A frame in the diffusion weight changes the inner product in which the bound is taken and leaves the bound in place. A frame placed elsewhere, in the matter coupling itself or as a cut-off below keV on the whole mean field, would defeat both arguments (Section~\ref{sec:shield}). \emph{What survives.} A classical theory of the Gaussian class must defeat both arguments at once, or leave the class. For noise made of Poisson jumps the floors and the mean number of photons stay as they are under stated conditions, and leaving the class helps only if nearly all of the keV photons arrive in multi-photon bursts of a kind the detector's event selection would reject (Section~\ref{sec:class}). The table of Section~\ref{sec:independence} sets the escapes side by side.