Mvuma cluster: four blocks, first read

A holder supplied four bare coordinates for a cluster of blocks he part-owns near Mvuma, Midlands: "a couple blocks we own, about 5 or 6", with crews working the ground now. The four points sit ~11 km WSW of Mvuma town, spanning 1.3 × 0.7 km. No claim polygons were supplied, so each coordinate gets a nominal 600 m square window (36 ha): whole-window numbers are neighbourhood statistics around each point, not claim statistics. A standard Zimbabwe gold block is ~10 ha, so each window covers its block wherever the point sits in it.

One provenance note worth recording: the coordinates arrived with no zone label, like every site in this study, but this time the holder's own GPS screenshot settled it. His app displays both zones at once, and converting its 35S readout to 36S reproduces the second readout within 8 m. The supplied eastings are WGS84 UTM 35S (EPSG:32735), confirmed, not just inferred. This ground also straddles the 30°E zone boundary on the data side: it is the first site in the study served by AlphaEarth's zone-36S tiles.

This is a working gold district. Four known producers sit within 40 km, and the closest is 1.7 to 2.6 km from every block (375 kg produced at 7.0 g/t). The Falcon group at Mvuma town, 11 km east, produced 12.7 t at 6.7 g/t plus 2.7 t at 4.8 g/t. A 169 kg producer at 16.6 g/t lies 6.4 km southwest, and it is the nearest deposit the regional model trained on.

The eastern pair carries the cluster: put the crews on blocks 3 and 2 first, block 4 last. Block 3 (860341 E / 7863302 N) is the strongest whole-block ground of the four: median P(gold) 0.325, the 82nd percentile of the local background, with the cluster's best spot, P 0.565 at 860600 E / 7863184 N, 290 m ESE of the supplied point. Block 2 (860550 E / 7863911 N) has the hottest supplied point itself, P 0.382, the 90th percentile: whatever is being worked at that exact spot, the model agrees with it, and its best spot (P 0.559 at 860420 E / 7864154 N) is a 280 m walk NNW. Block 1 reads mid-pack (median 0.268, 69th percentile). Block 4 (859289 E / 7863213 N) is the weak one: its point reads 0.175, the 37th percentile, below the local background median, and its best ground (P 0.514) actually sits between blocks 4 and 1 at 859560 E / 7863374 N.

For scale: block 3's whole-block read (0.325) sits just below the Silobela ground this same model scored at 0.346 to 0.359, the ground whose walk-first spot has now assayed 2.33, 1.52, 1.34 g/t on three of five grabs. It is above every Kwekwe block in the study. That is a resemblance statement, not a grade prediction: only rocks from these blocks can say what they run.

Every 10 m pixel over the cluster is inspectable: the heatmap is the fused P(gold) surface (dark = low, bright = high), and tapping anywhere shows the score, its rank against the local background and within the block, and every layer behind it, with the coordinate in peg format. The low ground is as visible as the high ground on purpose: control samples from dark pixels are what calibrate this model, so record those too. Scores read surface geology, not grade.

Block by block

Same estimate over each 600 m window; percentiles are against the shared 13 × 13 km local background. "P at point" is the model read at the exact supplied coordinate.

Block (supplied point, UTM 35S)	P at point	Window median	Best spot	Read
Block 3 · 860341 E / 7863302 N	0.329 (83rd)	0.325 (82nd)	0.565	strongest whole-block ground; walk it first
Block 2 · 860550 E / 7863911 N	0.382 (90th)	0.286 (74th)	0.559	hottest supplied point in the cluster
Block 1 · 859526 E / 7863562 N	0.280 (72nd)	0.268 (69th)	0.514	mid-pack, above local background
Block 4 · 859289 E / 7863213 N	0.175 (37th)	0.231 (57th)	0.514	weakest; its best ground lies toward block 1

The walking plan, one line per block:

Block 3: from the point, 290 m ESE to 860600 E / 7863184 N (P 0.565, the cluster's best), then the western line 860260 E / 7863244 N (0.499) to 860110 E / 7863304 N (0.498). One ~700 m hook covers the three.
Block 2: sample at the supplied point itself (0.382, 90th percentile), then 65 m NE to 860570 E / 7863974 N (0.521) and 280 m NNW to 860420 E / 7864154 N (0.559).
Blocks 1 and 4: one shared target between them, 859560 E / 7863374 N (P 0.514), 190 m south of point 1 and 280 m NE of point 4; block 4's own interior adds 859380 E / 7863024 N (0.476).

The five checks behind that

Check	This cluster	What it means
Regional P(gold)	union median 0.273 (70th percentile); blocks ranked 3 > 2 > 1 > 4	the combined estimate at 10 m; the eastern pair reads like the study's assayed Silobela ground, block 4 reads like background
Terrain match	97th percentile	high, but weakly calibrated here: only 4 known mines within 25 km to learn from (LOO AUC 0.68; spatial CV undefined), so treat as supporting, not primary
Red-oxide (ferric)	58th percentile	slightly above background on fully bare ground (0.02% vegetated, 0% cloud); no strong surface oxide anomaly, and too few local mines to test whether ferric marks gold here
Structure / lineaments	63rd percentile; dominant trend 105–120°	mildly elevated lineament density over the cluster with a consistent WNW-ESE grain; roads and field edges are the usual false positive
ASTER alteration (SWIR)	Al-OH 59th, carbonate/Mg-OH 62nd, chlorite 38th	archive 2006 scene, 0% cloud, ~1% vegetated: nothing anomalous either way

Compared to the other ground in this study

Site	P(gold) median	Percentile vs local background
Belingwe-West	0.394	n/a
Silobela site 2 window (assays up to 2.33 g/t nearby)	0.359	71st
Silobela block 3 (walk-first spot assayed 2.33 g/t max)	0.346	69th
Silobela site 1 window	0.331	63rd
Mvuma block 3 (this study)	0.325	82nd
Mvuma block 2	0.286	74th
Kwekwe block 3	0.285	36th
Mvuma block 1	0.268	69th
Kwekwe block 2	0.254	45th
Kwekwe block 4	0.238	34th
Mvuma block 4	0.231	57th
Kwekwe block 1	0.129	8th

Blocks 1 to 3 land in the same band as the Silobela ground, above all of Kwekwe. Note the percentile column reads high here partly because the Mvuma background itself is quieter than Silobela's: use the absolute P column for cross-site comparison and the percentile for within-site contrast.

Honesty notes, specific to this site

Edge of the training region. The regional model trained on 330 gold deposits and background sampled west of 30.39°E; these windows sit 2 to 4 km east of that edge. The nearest training positive is the 16.6 g/t producer 6.4 km southwest. The embedding generalizes smoothly and the surrounding ground is represented, but this is the first site scored at the training window's very edge, and the Falcon group itself was not in the training set.

Active workings. Crews are on this ground now. The AlphaEarth surface embedding is the 2024 annual mosaic, so it reads the surface as of 2024, before at least the current phase of work; but any ground already disturbed by 2024 (this is a historic district) can imprint the embedding, and the model is known to be disturbance-prone. The high read at block 2's point, where work is reportedly happening, should be read with that in mind: the model may partly be recognizing ground that miners already chose.

Data used

Regional P(gold) estimate: AlphaEarth surface data 2024, 10 m pixels (native, zone-36S tiles via Source Coop)
Red-oxide (ferric): Sentinel-2, scene 2025-08-20, 10 m pixels, 0% cloud, 0.02% vegetated
Hydrothermal alteration: ASTER SWIR band ratios, 30 m, archive scene 2006-06-11 at 0% cloud (SWIR failed in 2008; only archive scenes carry it)
Structure and terrain: Copernicus GLO-30 elevation, 30 m pixels
Known mines: Blenkinsop ZGS national gold-deposit compilation

The P(gold) grid is 10 m (~0.01 ha cells), so it pins the spot to walk, but it reads surface geology. It tells you where to look, not where a vein is. Rocks from the ranked spots, plus two or three from a low-scoring patch as a control, would calibrate all of it, exactly as the Silobela assays just did for that site.