Kwekwe block: a free-EO peg assessment

A second case study after the Zvishavane claim. A holder supplied a 157 ha block near Kwekwe in the Midlands greenstone belt, available to peg, chosen on the "red oxides, like the neighbouring mines" thesis, and asked how it performs on free satellite data. Kwekwe is a genuinely high-production district: there are 85 known gold mines within 40 km (the biggest neighbour produced 124 tonnes), and they are the local reference this analysis is built on.

Three independent folia-native methods, no Earth Engine: red-oxide (ferric) mapping, an explicit DEM terrain-signature match to the surrounding mines, and the validated AlphaEarth embedding analog. The map below carries all of them. The detail is in the Zvishavane method pages; this page is the verdict for the block.

Bottom line: free EO gives no peg-worthy signal here. The block sits in a favourable district and in mine-like terrain (right neighbourhood), but by the strongest validated method its surface signature looks like ordinary country rock, not gold ground (49th percentile, best corner well below the gold band). And the red-oxide rationale is not corroborated: the block is not iron-anomalous, and the nearby mines are only weakly ferric-elevated in the imagery, so ferric staining cannot rank the block at 10 m. This is prioritization evidence against, not proof of barren ground (satellites miss sub-pixel veins and in-pit oxides). The spots below are best-case corners to walk first, never gold.

The block spots, as a list (UTM 35S, the holder's grid):

What the map shows

Embedding P(gold), 80 m (on by default) is the validated method: how much the ground resembles known gold sites in AlphaEarth's 64-D satellite embedding, scored by a RandomForest trained on the 85 local mines. Over the block it is dim, the block reads like background. This is the layer to trust.
Red-oxide ferric ratio (S2 B04/B02) is the holder's thesis layer. It is not anomalous over the block, and (the test that matters) the known mines are only weakly ferric (nearest-pixel +0.26σ, 60th pct), so a bright patch here is weak evidence at best.
Gold mines sized by production. The block sits among real producers: the district is favourable. That is the strongest thing in the block's favour, and it is about the neighbourhood, not the block.
Spots are each method's best-case ground inside the block. They cluster in the north / north-west corner, which is the only constructive read here. None is a gold indication; all are walk-first-with-low-expectations points.

How this was made, and how it was validated

Each method trains on the 85 local mines and reports leave-one-out skill and a spatial cross-validation, the test for whether the skill is real or just spatial autocorrelation (clustered mines memorised by the model):

Method	Leave-one-out	Spatial CV (8 km buffer)	Block verdict
Embedding analog	AUC 0.94	0.90 (holds, strong)	49th percentile, background-like
DEM terrain	AUC 0.91	0.87 (holds, coarse)	85th percentile (mine-like terrain type)
Red oxide	mines weakly ferric (+0.26σ pixel)	n/a	not anomalous (40th percentile)

The embedding skill holds up strongly under spatial cross-validation, so the block's background-like score is trustworthy, not an artefact. The terrain signal is real but coarse (the mines sit on slightly more dissected greenstone terrain than the granite plateau, a "right terrain type" filter, necessary not sufficient). Recipes: research/mineral-prospecting/kwekwe-block/ (embedding.py, dem_terrain.py, iron_oxide.py).

Soundness check. Because known mines are physically disturbed ground (pits, tailings) and AlphaEarth embeds surface appearance, a ring test checked whether the model scores the disturbed footprint or real geology: P(gold) around the mines decays from 0.30 at the pit to 0.10 at 1.5 km but stays 5 to 7x background throughout, so the signal is largely areal greenstone-belt geology, not a mining artefact (a modest disturbance component remains). The block scores below even that 1.5 km mine-halo, so its background-like reading is not just an artefact of the block being un-pegged. The coordinate decode was independently validated against Globe & Phoenix and Cam & Motor (the latter's grade matches the published value exactly). Recipe: research/mineral-prospecting/kwekwe-block/validate.py.

The single biggest lever from here is one real ground observation from the block or a neighbour: it wires straight into the iron-oxide calibration and the embedding reference, and re-weights every layer against real data.