How the models work, and what they can and cannot tell you
Most prospecting maps show you where gold was found before. Pannable additionally simulates the physics that moves placer gold, on real terrain, and paints the result. This page explains each model in plain English, what the colours mean, and what we found when we tested them.
Not a guarantee of gold. Model output is a set of relative terrain scores: decision support for where to look first, not a measurement of gold and not a prediction that you will find any.
Informational only — not legal advice, and not a substitute for the advice of an attorney. Pannable synthesizes public land and claims records (BLM MLRS, PAD-US, USGS, state agencies) that can be incomplete, delayed, or wrong.
Verify before you dig: confirm current claim status on BLM MLRS and land-use rules with the managing office (BLM/USFS/state) at the actual site, every time. Prospecting on an active mining claim without the claimant's permission can expose you to civil liability and other penalties.
Physics, correct by construction
Placer gold moves only while water has the energy to carry it, and drops where that energy collapses. The models simulate that on the real ground: flow directions and accumulation are computed over the digital elevation model, stream power is derived from them, and each model scores a different part of the story. We do not curve-fit the models to find reports. Every constant is documented with its physical reasoning, marked where it is a judgement call, and adjustable, so the models improve by adding physics, not by chasing a dataset.
The foundation is terrain. Pannable bakes its own hillshade and models from raw USGS 3DEP 1 m LiDAR wherever it exists (the raw cloud-optimized GeoTIFFs, not the smoothed image service), and uses 10 m national elevation elsewhere. A missing 1 m source falls back to honest 10 m, never to a smoothed substitute, and every tile says which one you are looking at. The baked models run in roughly 1.6 km cells with a 500 m halo so flow does not stop at cell edges, and are served as pre-rendered tiles, so the map has nothing left to compute at view time.
What the colours mean
Each model paints five tiers, Very Low to Very High. They are relative terrain scores: a ranking of ground within the model’s own population, not a probability and not a quantity of gold. “Very High” means “among the best-fitting terrain for this mechanism,” and is the place to look first. It is decision support, not a result.
- Scores are pixel-crisp at 1 m; zoom in to around zoom 14 before the detail renders.
- Man-made ground is masked: road cuts, fills and embankments are blanked rather than scored, because the models do not simulate engineered terrain.
- A gold coverage outline shows where the 1 m models exist. Outside it, 10 m models cover almost anywhere with recorded mines.
- Mine-sourced models (Gravity, Composite) are legitimately sparse where there are no recorded lode sources: no source, no transport to simulate.
Five questions asked of the terrain
Gravity
Where has lode-shed gold travelled? Starts from every recorded hard-rock source and routes downhill transport along the real drainage network (flow direction and accumulation, not straight-line distance), decaying with distance. Needs a known mine; reach is on the order of a few kilometres.
Hydraulic
Where does the water drop its load? Computes stream power (a standard transport-capacity proxy from accumulation and slope) at every cell, then takes its derivative along the downstream flow path. The score is positive only where capacity is falling: the energy collapse that lets heavy particles settle. Needs no mine; works on the trunk network.
Composite default
Where do supply and settling agree? The geometric mean of Gravity and Hydraulic, so a hot zone needs gold arriving and gold being forced to stop, and disagreement is heavily penalised. Composite is also flood-aware: where recent high water has reworked a reach, that freshness is folded in.
Crevice
Can this bedrock hold gold? The finest-grained model, for the person standing on bedrock. On 1 m LiDAR it looks for exposed-bedrock micro-relief in and beside active channels (transverse ribs, potholes, crevices) and scores them as traps against fixed physical thresholds, so a featureless gravel valley scores nothing rather than “the best of a bad lot.”
Gulch
Which gully is delivering coarse gold, and where does it stop? The trunk-stream models gate on drainage area and so skip the small feeder drainages where coarse gold is shed. The gulch model covers that band and the transition at the gulch mouth. It is on the map at 1 m, only near recorded lode sources: with no source it scores nothing by design. Its constants come from spot checks and are not fitted to any district, and it was not part of the July 2026 backtests below.
| Model | Question | Needs a mine? | Flow-routed? | Scale |
|---|---|---|---|---|
| Gravity | Where has lode-shed gold travelled? | Yes (sources) | Yes | ~5 km reach |
| Hydraulic | Where does the water drop its load? | No | Yes | Trunk network |
| Composite | Where do supply and settling agree? | Yes (via Gravity) | Yes | Same as inputs |
| Crevice | Can this bedrock hold gold? | No | No | Metre-scale, local |
| Gulch | Which gully is delivering, and where does it stop? | Yes (tight) | Yes | Feeder band |
Flood, season phase and the River Reset Clock
The static models answer where. Three weekly-refreshed layers answer when, behind a freshness gate that refuses to publish stale “current conditions.”
Flood model
Rivers glow where live USGS gages show high water, with week, month and season windows and a flood-accumulation heatmap. It only lights up near gages, so it is sparse where gages are sparse, and the map says so.
Season phase
Each model district is classified from SNOTEL snowpack and streamflow: pre-freshet (scout, don’t dig), freshet (gold moving, banks dangerous), recession (fresh-worked gravel exposed), baseflow (crevicing and in-stream work). Eastern districts have no SNOTEL and use gage-only rules.
River Reset Clock
For every gold-country reach, when did the last armor-breaking flow happen? Per-reach thresholds come from 44 years of hourly National Water Model history; current conditions are compared against them weekly and reaches classify as stable, transporting, fresh or reworked.
The backtests, including the parts that do not flatter us
The models are not fitted to find reports, but we did run two adversarial tests in July 2026 so that nobody has to take the physics on faith.
- Held-out lode mines at 1 m. Simple proximity to known mines captured held-out lode occurrences about 2.9× better than chance; the best physics model (Gravity) managed 1.7×; Composite was below chance. This test is structurally mismatched (the models score the modern drainage, while a hard-rock mine sits on the hillside vein), so it is a lower bound, not the intended target.
- Held-out placer occurrences at 10 m, the fair target. Across 315 held-out placer occurrences in 12 Sierra Nevada districts, Gravity and Composite beat chance by a real but modest margin (lift about 1.3×), and a distance check confirmed the signal is not noise. But starting near known placer workings beat every model by roughly three to one (lift about 4×), and even proximity to known lode mines beat all three physics models.
The honest conclusion: the physics is a legitimate additive input, not a standalone predictor that out-ranks “go where gold has been found.” That is why the Top Spots ranking weights historical evidence at 55% and trap physics at 25%, and why the mines layer is always one click away from any model. Treat every model tier as a physics-ranked candidate, not a result.
What the models do not know
- They score terrain, not gold. A perfect trap with no source upstream holds nothing.
- They know nothing about what has already been worked out. A hundred and fifty years of dredging and hydraulicking is not in the elevation model, except where it changed the ground.
- Coverage is uneven. 1 m models cover most western mining districts and the southeastern gold belt, not Alaska and not everywhere; 10 m models are coarser and smoother.
- On-demand model runs for an area you draw are capped by compute time, so large areas are tiled or queued.
- None of this is a legal reading. A high score on claimed ground is still claimed ground.
Verify before you dig: confirm current claim status on BLM MLRS and land-use rules with the managing office (BLM/USFS/state) at the actual site, every time. Prospecting on an active mining claim without the claimant's permission can expose you to civil liability and other penalties.