Backdoor-Trigger Evaluation for Computer Agriculture Disease-Detection Models Under Third-Party Pretrained Checkpoints: An Artificial Intelligence Poisoning-Rate Audit

Third-party pretrained checkpoints are now the default starting point for computer-vision disease detectors in digital agriculture, yet the poisoning rate at which a backdoor trigger survives fine-tuning on clean crop imagery has not been characterised as a design quantity. the present analysis treats the backdoor trigger as a property of the checkpoint supply chain rather than of the attacker, and asks what poisoning rate a defender must assume when the provenance of the pretrained weights is unknown. A formal model of trigger persistence is developed in which the backdoor is represented as a parameter-space displacement whose survival under gradient fine-tuning depends on the ratio of trigger-aligned to task-aligned curvature. From this model a persistence bound is derived, and the bound is then embedded in a before-after-control-impact design that treats the poisoned checkpoint as the intervention and the clean checkpoint as the control. Operating characteristics were evaluated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, under a declared variance structure with between-cluster SD of 6.0, residual SD of 4.5 and intraclass correlation of 0.64. The design required 54 clusters to detect a 5 percentage-point difference at 80% power.