Virtual-world physics engines cannot hold a graviton: the correction only becomes storable beyond lunar orbit, while AI low-precision formats keep the signal

The static gravitational correction implied by a massive graviton falls below the smallest difference any floating-point format can hold at every scale virtual worlds use, so physics engines cannot store it. Whether graviton physics could be represented inside the computational systems that run virtual worlds and AI had not been tested; the only quantified property was an upper bound on the graviton mass. Using a 60-digit reference over 301 log-spaced distances from 1 m to 10¹⁵ m, the correction first becomes storable at 5.40×10⁸ m in float64 and 1.25×10¹³ m, about 84 AU, in float32 at the published bound.

Evaluating the Yukawa correction naively, rounding after every operation, produces phantom corrections where the true value cannot be represented, with the same sign as the real effect. Invisibility would be the benign outcome, but cancellation error turns it into fabrication. In float64, 58 of 175 grid points where the true correction is unrepresentable return a non-zero value, up to 9.2×10¹⁶ times the true one; in float32, 22 of 262 points do, up to 1.3×10⁸ times.

The same graviton physics survives in the low-precision formats used by AI hardware, so precision is not the binding constraint there. Low-precision rounding was expected to erase the dispersion residual at the bound, giving a representational SNR below 1. After time-and-phase alignment, representational SNR ranges from 9.7 for FP8 E5M2 to 2.0×10⁷ for float32; a discriminator trained on noise-free waveforms reaches AUC 1.000 in every format above the bound, with no degradation as precision falls.

The conclusions cover the static correction in linearised Yukawa form and a single GW150914-like inspiral waveform restricted to the inspiral stage. Vainshtein screening in non-linear massive gravity would make the static correction smaller still, so the thresholds are conservative in the direction of visibility. The discriminator experiment uses synthetic waveforms without detector noise, and the network computes in float32, so it tests storage precision of inputs rather than low-precision arithmetic inside the network. The intended readers are developers of virtual-world engines, digital twins and gravitational-wave machine-learning pipelines.

Raw strain at its astrophysical amplitude, peak 10^-21, is flushed entirely to zero in float16 and both FP8 encodings, while bfloat16 stores it with 0.16% relative error; the favourable result holds only after the normalisation machine-learning pipelines routinely apply. At the preprint bound and at 6×10^-24 eV/c2 the discriminator is at chance in every format even though the residual remains representable there, so the limit is learnability under nuisance phase and time. The bound itself differs between the preprint and published versions of the same analysis; both were carried through and the conclusions do not depend on the choice.