Embedding-analog mineral prospectivity — Mberengwa belt, Zimbabwe

validated

Does a place's 64-D AlphaEarth embedding resemble the embeddings at a belt's known-mineralized sites more than it resembles random country rock? What is graded is the held-out skill of that analog on this belt, measured leave-one-out in the same run that draws the map. That skill is a BAND, AUC 0.92 to 0.96 across background draw and sampling grid, not a single figure, and it is Mberengwa-specific: the reference script reads 0.9608 and the engine recipe a visitor can re-run today reads 0.9253. The site read itself is a resemblance diagnostic with no ground truth, never a detection.

Methods

embedding analog random forest leave-one-out validation

Registry ops (5)

embedding_field raster_sample_points point_classifier_loo_validate point_classifier_train point_classifier_predict

Recipe source: registry/recipes/folia/reproductions/embedding-analog-gold-prospectivity/folia.yaml — view folia.yaml

Reference: research/mineral-prospecting/occurrences/occurrences.geojson

metric	observed	threshold	measured by
heldout_positive_vs_background auc	96.1%	≥ 0.75	human · 2026-06-24
median_heldout_percentile overall accuracy	99.1	≥ 75	human · 2026-06-24
heldout_above_90th_pct producers accuracy	89.4%	≥ 0.6	human · 2026-06-24
site_resemblance_read note	pending	≥ 0
heldout_positive_vs_background (engine) auc	92.5%	≥ 0.75	agent · 2026-08-19
median_heldout_percentile (engine) overall accuracy	96.7	≥ 75	agent · 2026-08-19
heldout_above_90th_pct (engine) producers accuracy	76.5%	≥ 0.6	agent · 2026-08-19

Validation history

validated 2026-08-19

A scheduled validation agent re-runs each reproduction and flips the badge on drift (hub Phase 1).

Fork this reproduction

view folia.yaml view on github