Candidate availability and uncertainty-aware selection shape the reliability of dairy rations under variable feed composition

Abstract Least-cost dairy rations meeting nutrient limits at mean feed composition can fail when composition varies. We compared fixed and nutrient-informed selection rules on matched candidate pools for one reference cow using selected NASEM equations, with actual dry matter (DM) hidden during selection. With planned DM increased by 10%, adding admissible candidates reduced failure in every comparison; compression for training coverage increased it in five of six. Full-pool selection reversed these losses in all 18 test comparisons, with risks no greater than 2.155% without cottonseed and 0.825% with it. An exploratory least-cost linear program re-solved at the observed composition with the best of four DM margins failed in 8.1--9.8% of states. At higher simulated measurement noise, using posterior means alone increased failure, while adding posterior variance to the score reduced it. Energy violations became more frequent but smaller on average when they occurred. With a 20% higher microbial-protein coefficient, full-pool risks for this uncertainty-aware rule exceeded both case-specific targets, although its high-noise advantages remained. Reference-supply cases remained high-risk. These findings assume consumption of all supplied DM, use an illustrative crude-protein target and have not been field-validated.