Evaluation of Attention-Rollout Consistency Audits for Computer Vision Agriculture Weed-Detection Models Under Variable Tillage Residue: An Interpretability Localization Bound for Mechanical Weeding Safety

Attention-rollout consistency audits are increasingly proposed as a safety check for vision-based weed detection in mechanical weeding, yet the statistical behaviour of the audit's localization bound under variable tillage residue has not been characterised. the present analysis formalises the audit as a bounded-error estimator of the spatial correspondence between rollout attention mass and the agronomic target region, and derives a finite-sample localization bound under a declared data-generating process. A replication-based evaluation was conducted with 600 data sets of N = 500 drawn from a process with intercept 12.0, exposure 0.42, group -3.1, severity 1.35 and error SD 4.2, using seed 1424973492. The ordinary least squares estimator was characterised across replications. Estimated bias was 0.03246 for the intercept and -0.00024 for exposure; root mean squared error was 0.82788 for the intercept and 0.01517 for exposure; empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters. The results indicate that the audit's localization bound is conservative under the stated process, and that residue-induced attention drift is absorbed by the bound rather than by the point estimate. Rollout consistency audits should be reported with their coverage properties, not only with pointwise attention maps, if they are to function as safety evidence for mechanical weeding.