One acceleration scale per galaxy? Gas and stellar dynamics disagree inside star-forming discs at z ≈ 0.8

We ask whether one acceleration scale describes the radial acceleration relation (RAR) at two radii inside the same high-redshift galaxies. LEGA-C star-forming discs at z ≈ 0.8 provide ionized-gas kinematics and stellar Jeans-model (JAM) masses at the effective radius R_e and at 2R_e. With a fixed baryon model and the published uniform-dispersion pressure correction, the RAR scale fitted to the gas rises from a0 = 0.71 at R_e to 1.43 at 2R_e (10^-10 m s^-2) in the 40 galaxies flagged as having extended data: Δ = log a0(2R_e) − log a0(R_e) = +0.30 dex (95% +0.15 to +1.01), and +0.13 (+0.08, +0.19) in all 147. With JAM masses of 38 of the same galaxies, Δ = −0.08 (−0.24, +0.05), and the gas–JAM difference in Δ excludes zero. The per-galaxy JAM–gas mass offset falls outward, by −0.16 (−0.20, −0.12) dex on average; its sign is robust (35 of 38 galaxies), and its size tracks the JAM precision. The same two-radius fit on 138 local SPARC discs with their measured mass models gives Δ = −0.02 (−0.07, +0.04). With the LEGA-C stellar-disc recipe it gives −0.02 on the 53 massive discs but +0.14 (+0.09, +0.20) on all 138, which are mostly low-mass discs. Gas and stellar dynamical-mass profiles therefore disagree between R_e and 2R_e, as the published aggregates on these galaxies already imply. For the RAR, the consequence is a radius-dependent difference between the gas and stellar a0, amplified three- to six-fold by the position of these galaxies on the relation. If the JAM profiles are right, the rise sits in the inner gas kinematics. In exploratory refits, published alternative pressure-support prescriptions leave a rise of +0.19 to +0.30 dex. These data do not constrain the fitted a0 levels: a0(2R_e) = 1.43 (0.84–2.28) contains the local value, a horizon-scaled value and the MUSE-DARK model at z = 0.8, and the JAM fits imply levels of 5–9. All four analyses were preregistered and audited by independent AI sessions before a single scoring run, with self-recorded locks; the trail is public (a0(z) record v1.6, doi:10.5281/zenodo.23444426).