View the recipe

This is the file that runs. Not a description of the method, not pseudocode for it: the governed spec the embedding-analog report is computed from, reproduced here verbatim from the repository at build time. Every number on that page comes out of running it.

folia run registry/recipes/folia/reproductions/embedding-analog-gold-prospectivity --no-cache

That one command is the whole instruction. The AlphaEarth reads happen inside the run: a column-projected lookup in the published tile index over HTTPS range requests, a windowed read of each intersecting Cloud-Optimized GeoTIFF, int8 dequantisation per the dataset record's declared encoding, and a mosaic across UTM zones onto the analysis grid. No prefetch step, no Earth Engine, no credentials, nothing materialised on disk first.

Why "view the recipe" and not "view the code". The sampling, the background draw, the forest, the leave-one-out harness and the scoring pass are operations in a shared engine, named and versioned. What is specific to this study is entirely in the spec below: which belt, which reference points, which window, which seeds. That is the part you can read in a few minutes and disagree with, which is the point of publishing it.

What makes it a different site

Belt, site, reference set and every hyperparameter are parameters. Standing up a second site is a --set list, not a fork of the recipe:

Parameter	What it is
`belt_bbox`	The belt the model learns "deposit vs country rock" within, and where the background is drawn from.
`site_bbox`	The AOI you want a resemblance read on. It has to sit inside the belt.
`occurrences_uri`	Your belt's known-occurrence GeoJSON, travelling with the recipe (`local://data/…`) or fetched over http(s).
`positive_filter`	Which features count as positives. Non-matching features stay in the file and still push the background draw away from near-deposit ground. `{}` = every feature is a positive.
`background_seed`, `rf_*`, `reference_window_px`	The knobs the measured skill was measured under. Change them only with a reason, and re-measure.

The fill-in-the-blanks form of that list, a template, and the launcher that writes the command live in the repository at research/mineral-prospecting/embedding-analog/sites/. A second site, the Midlands (Kwekwe-Gweru) belt scoring Kwekwe block 4, is staged there as the worked example.

What a new belt owes. Held-out skill does not travel. The Mberengwa belt measures leave-one-out AUC 0.92-0.96, where the spread is the background draw and the sampling grid, against a 0.75 floor. A different belt inherits none of that: it re-runs leave-one-out against its own reference set and earns its own number, or its site read is uninformative in either direction. The floor is what transfers, because it is a statement about what counts as usable skill rather than a measurement.

The spec

Where the rest of it lives

The spec above cites three things by name, and all three are readable:

The dataset record. @google-deepmind/aef/v1/annual declares where the tile index sits, which columns it carries, and how int8 values dequantise. That record is the contract; swapping embedding products means swapping the reference, not editing the engine.
The operations. raster_sample_points, point_classifier_train / _loo_validate / _predict, and the embedding_field source arm. None of them knows anything about gold: the chain is "a multi-band raster plus labelled points, in; a classifier, its held-out skill, and a score map, out."
The validation record. validations/comparison.yaml states the floors first and then the measurements, and decomposes the 0.92-0.96 band by swapping one thing at a time (the classifier is worth ~0.001-0.01, the particular background draw ~0.02, the reprojected sampling grid ~0.015).

The output is a resemblance diagnostic, not detection. A high score means this ground's multi-temporal optical, radar and elevation signature, as compressed by AlphaEarth, looks like the ground at known deposits. AlphaEarth embeds surface appearance and seasonality, not subsurface mineralisation. The reference set is ~95% one source and records discovered, historically produced deposits, so the model can only learn the look of the deposits we already know about.