Photographs interpreted by machine vision on above-ground gas pipelines vary in sharpness, exposure and compression because of how they are acquired, and that variation degrades network performance. One safeguard is an image-quality gate: Admit photographs passing a criterion, route the rest to manual review. This study asks whether such a criterion can be calibrated to a stated target, what share it admits, and what is withheld. A gate combining a luminance-normalised sharpness statistic with an exposure envelope was calibrated on the validation subset of a family-partitioned corpus of 9,965 pipeline image records, using five instance-segmentation checkpoints as the instrument. The frozen, hashed rule was applied once to a 150-image evaluation subset under fourteen conditions. For deformation, the primary class, no acceptance zone was established: The estimate never reached the target, so the threshold was infinite and the primary endpoint, the change in residual failure among accepted photographs, is not estimable. Two secondary rules accepted 42.7% and 47.2% of image versions, with estimated residual-failure differences of −0.0226 and −0.0248 against the same stream without gating, and intervals including zero; a reduction was not demonstrated. Against nominal performance on unimpaired photographs, a separate comparison point, non-inferiority within the pre-specified margin was also not demonstrated. The quantile-based fallback rule would leave 59.3% of deformation object–seed observations outside automated interpretation, 20.2% of them already correctly matched: Implied workload, not measured cost. The study is a retrospective internal evaluation with a self-documented freeze; deviations found after the evaluation subset had been examined are disclosed.