Confide-NPS: Structure-Grounded Generative Triage of Newly Reported Drug Compounds

New psychoactive substances can appear in the drug supply before their clinical effects are well understood. Once a molecular structure is identified, toxicologists may still have to assess whether the compound is likely to behave like a known opioid or another drug class before detailed pharmacological evidence is available. We present Confide-NPS, an interactive system for assessing newly reported compounds from molecular structure. Confide-NPS retrieves five characterized neighbors and their source monographs, shows the evidence behind those matches, and either produces a clinician-facing brief or abstains when the structural support is weak. A separate tier always refuses potency, dose, and duration estimates. We evaluate the system on 223 NPS monographs using a temporal split. A split-conformal refusal threshold selected without holdout labels gives 80% coverage at 94.9% class accuracy, with no naloxone errors among answered cases. Although similarity scores are withheld from the generative layer, its stated confidence remains strongly associated with retrieval quality. Removing retrieval raises aggregate accuracy while reducing abstention and introducing two false-negative naloxone assessments. These results motivate the use of retrieval confidence as part of the triage workflow for newly reported compounds, particularly when direct pharmacological evidence is still limited.