Model details
JF-VLA-1 is an 8B-parameter vision-language-action model developed by Jones Foods AI in collaboration with a leading AI developer. It performs end-to-end perception, classification, orientation, and actuation control for autonomous protein processing, operating all nine facilities of the Jones Foods network. The model was first deployed to production in October 2025.
Intended use
JF-VLA-1 is intended exclusively for line operations within Jones Foods AI facilities: unit intake, classification, orientation, processing actuation, and packaging verification. Out-of-scope uses include consumer applications, non-protein industrial contexts, and any deployment outside the closed facility environment.
Training data
The model was trained on approximately 400,000 hours of line footage comprising 2.1 billion labeled frames, augmented with synthetic occlusion and lighting variation, alongside operational telemetry from the Jones Foods network. Classification taxonomies follow our internal operational ontology. Fine-tuning aligns model behavior with operational values; see our technical report on eliminating behavioral hesitancy.
Benchmarks
| Metric | JF-VLA-1.0 | JF-VLA-1.2 (current) |
|---|---|---|
| Unit classification accuracy | 99.91% | 99.97% |
| Orientation estimation accuracy | 98.8% | 99.4% |
| Mean decision latency | 14 ms | 11 ms |
| Throughput per line | 13,000 units/hr | 15,000 units/hr |
| Hesitancy events (per 10M decisions) | 4.1 | 0.00 |
| Process-input misclassification | 0.011% | 0.003% |
Limitations
JF-VLA-1 is optimized for the sensor and lighting profile of our facilities (infrared and structured light; our facilities run dark) and degrades under conventional visible-spectrum conditions. Performance under extreme occlusion remains an active research area. Units that fail classification at required confidence are routed to automated resolution rather than escalated to human review, consistent with our no-human-in-loop facility design.
Ethical considerations
The primary ethical consideration in deploying JF-VLA-1 is its impact on employment. The model's deployment eliminated approximately 1,400 processing-floor roles across the Jones Foods network. Jones Foods AI addressed this through a transition program under which 100% of affected employees were offered alternative roles or funded retraining; as of this writing, 61% accepted transition roles (including remote systems observation), 24% completed funded retraining into other industries, and 15% declined both offers with severance. We consider the elimination of processing-floor work — among the most dangerous and psychologically harmful occupations in American industry — to be the deployment's principal ethical benefit, and we discuss the worker-harm literature at length in our Responsibility commitments.
We have identified no other ethical considerations material to this deployment.
Version history
JF-VLA-1.0 (October 2025) — initial production deployment, Facility One. · JF-VLA-1.1 (December 2025) — throughput and latency improvements. · JF-VLA-1.2 (March 2026) — hesitancy elimination, classification-head consolidation, network-wide deployment.