From minutes to bounds: A probabilistic UV-C control and a shape-only morphological fingerprint for postharvest Colletotrichum

Postharvest losses in high-value horticultural crops such as cacao and coffee are often driven by Colletotrichum spp. and other latent fruit pathogens. Ultraviolet-C (UV-C) is increasingly deployed as a chemical-free postharvest technology, yet prescriptions remain framed in minutes rather than in probabilistic guarantees of disease control and safety. We present a dual framework that (1) establishes conservative confidence bounds on survival and (2) validates a “shape-only” morphological fingerprint. This approach addresses the biological complexity of host-pathogen interactions by quantifying isolate-specific heterogeneity, rather than averaging it away. We utilized a large Colletotrichum dataset (n = 5,363) from cacao and coffee, spanning diverse treatments including UV-C, UV-B, and sonication. First, focusing on the Coffee UV-C cohort (∼10 min), we quantified this heterogeneity; the most conservative Clopper–Pearson upper 95% bound on survival reached 1.000, highlighting partial survival events (e.g., isolate P24-88) under otherwise high-efficacy conditions. This probabilistic framework captures the “tail-risk” of biological resilience instead of assuming complete kill. Second, we trained machine learning models on the full dataset using only geometric features (e.g., aspect ratio, asymmetry), explicitly excluding all primary size metrics. Serving as a rapid physiological indicator of UV-induced stress, these “shape-only” models successfully predicted pathogen host-origin (Accuracy ≈ 0.93) and post-treatment survival (R² ≈ 0.74). The signal’s ability to generalize across UV-B and sonication confirms that geometry, not just growth reduction, carries a robust and transferable physiological stress signature. This work provides a device-agnostic, probabilistic control platform, replacing time-based heuristics with quantitative guarantees and a generalizable, shape-based diagnostic. A conservative upper 95 % confidence bound (on survival = 1.000) was observed for the Coffee UV-C (∼10 min) cohort, revealing strong isolate-specific heterogeneity rather than a universal kill guarantee Isolate-specific heterogeneity (e.g., localized P24-88 survival) was quantified rather than averaged away. A size-free morphological fingerprint predicted host-origin (Accuracy ≈ 0.93) and survival (R² ≈ 0.74). The shape-only signal generalized across diverse stressors, including UV-C, UV-B, and sonication. A probabilistic, device-agnostic control framework replaces traditional time-based heuristics. A conservative upper 95 % confidence bound (on survival = 1.000) was observed for the Coffee UV-C (∼10 min) cohort, revealing strong isolate-specific heterogeneity rather than a universal kill guarantee Isolate-specific heterogeneity (e.g., localized P24-88 survival) was quantified rather than averaged away. A size-free morphological fingerprint predicted host-origin (Accuracy ≈ 0.93) and survival (R² ≈ 0.74). The shape-only signal generalized across diverse stressors, including UV-C, UV-B, and sonication. A probabilistic, device-agnostic control framework replaces traditional time-based heuristics.