Pezego-HITL: A policy-grounded large language model architecture for agricultural extension in Ghana

Policy alignment is critical for high-stakes AI in agriculture: The paper demonstrates that standard AI quality metrics are insufficient for agricultural decision-support systems. The researchers developed P-EVAL, a specialized evaluation framework that measures not just accuracy but also "Policy Alignment Rate" (PAR of 0.94), ensuring AI recommendations comply with official agricultural guidelines—a key requirement for smallholder farming contexts where incorrect advice can have serious consequences.

Human experts remain essential in the loop: The Pezego-HITL architecture incorporates human-in-the-loop (HITL) design, meaning expert agronomists review and validate AI recommendations rather than fully automating decisions. This approach acknowledges that AI alone cannot be trusted for high-stakes crop protection decisions and that expert supervision workload must be actively managed as part of responsible deployment.

Speed and accessibility matter for real-world impact: The system reduced response time by 55% (from 28.6 to 12.9 seconds) through intelligent caching, achieving a 59.6% cache reuse ratio. For extension workers in Ghana serving multiple farmers, this latency reduction directly affects whether the tool is actually usable in field conditions, demonstrating that responsible AI requires attention to practical operational constraints, not just accuracy.

Smaller, open-source models can be effective alternatives: The researchers showed that a smaller open-source model (Qwen3.5-9B) achieved strong performance (PAR of 0.86) with even faster response times (10.2 seconds), suggesting that responsible AI deployment doesn't require the largest proprietary models and can instead leverage more transparent, auditable alternatives.

Socio-technical integration requires direct stakeholder input: The evaluation included "detailed questionnaires to Ghanaian Extension Services," indicating that the researchers recognized responsible deployment demands understanding how local