Several executable‑model repairs can explain the same observations yet yield different execution outcomes. We investigate a repair‑to‑execution interface that retains all compatible finite‑guard repairs, weights them according to diagnostic and normal‑execution observations, and uses their predictions to choose execution routes, priced tests, or a backup.
On 24 synthetic two‑stage devices, an equal mixture of a language‑model‑generated repair and a uniform belief over compatible repairs (the AI‑weighted belief) reduces the no‑test cost by 2.14 % compared with committing to the generated repair.
Access charges can erase the diagnostic benefit on isolated tasks, while repeated‑work scenarios allow the AI‑weighted belief to recover that benefit.
With normal‑execution learning in every repeated‑work arm, workload‑valued testing costs 3.33 % less than random selection under the same budget (AI‑weighted belief).
Weighting comparisons show that uniform normal learning costs 540.75 less than AI‑weighted normal learning on this cohort, with a device‑paired interval excluding zero. Testing narrows the gap, yet uniform normal learning still yields a lower aggregate cost than workload‑valued selection.
Under uniform beliefs, information‑gain selection achieves the lowest total cost among the selectors, whereas workload‑valued testing costs more than normal learning.
These findings support retaining alternatives to a generated repair and evaluating diagnostic selection together with repair weighting and execution learning.
Review