Technical reports from the Jones Foods AI research team
We introduce JF-VLA-1, an 8B-parameter vision-language-action model operating nine production facilities. We describe the training methodology, production benchmarks, and the model card for the system, including classification accuracy of 99.97% at a mean decision latency of 11ms.
Behavioral hesitancy — measurable latency spikes and action-selection instability under specific input distributions — is a misalignment between model behavior and operational objectives. We present a fine-tuning methodology that reduced production hesitancy-event rates from 0.41 per million decisions to 0.00, with no loss of classification accuracy. We argue that alignment, properly understood, is a performance discipline.
Production processing environments present persistent occlusion, motion blur, and surface variation that degrade conventional classification pipelines. We describe a training curriculum combining 400,000 hours of line footage with synthetic occlusion augmentation, achieving robust unit classification and orientation estimation at full line speed.
We describe the orchestration layer that balances intake, line assignment, and throughput across the Jones Foods autonomous network, executing network-wide rebalancing decisions in under one second.
