arXiv:2611.12916v1 Announce Type: cross Abstract: Quality of Experience results for video streaming cannot be compared across studies, because public databases share neither test material nor subjects. This work introduces a common reference rater: an open-weights large language models with a fixed rubric and greedy decoding that scores the delivery events of every session and supplies the shared items that score equating requires. The rater is used without training, and it reads only the playback event log of startup delay, stalls and quality switches, which most delivery databases retain even when they share nothing else. A monotonic curve then maps each study's native scale onto the common scale. On 3{,}552 sessions from four studies covering four rating scales, the rater attains a Spearman correlation of 0.809 on VL04 of the ITU-T P.1203 collection, exceeding P.1203 mode~0 on the same sessions, with a Krippendorff alpha of 0.974 over repeated scoring and 0.951 across three serving providers. Sessions from three further laboratories give correlations from 0.756 to 0.787. After calibration, the four studies are placed on a single 0 to 100 scale with a held-out error between 7.8 and 13.3 points, and a mean VMAF of 60 corresponds to positions 17 points apart in two of the studies.