Relief honesty: these layers are visualization-grade shaded relief (Esri Ocean / GEBCO). Good enough to learn the workflow and flag coarse anomalies; wreck-scale detection runs on the BlueTopo survey pipeline — next phase. Every candidate you file carries the layer it was flagged on.

Anomaly Bench

Train on the found · scan for the unfound

Live · Lake Superior

Reaching NOAA…

Relief layers

GEBCO shaded relief Known wrecks Filed candidates

Paste a tile template from nowCOAST (BlueTopo WMTS) or any XYZ source. Saved locally; used as the scan surface when present.

0 · Synthetic pretrain

Synthetic pos / neg0 / 0

Conditioned ondefaults

Procedural relief patches whose roughness follows the live sea state off the NOAA buoys; positives carry a hull-scale ridge with a hillshade shadow pair. Feeds the training pool, and exports as a labeled dataset with full provenance.

1 · Collect ground truth

Positives (known wrecks in view)0

Negatives (empty lakebed)0

Held out for validation0

Zoom to a wreck field (z ≥ 9), let tiles load, collect. Move, repeat — the dataset accumulates across views. A quarter of the wrecks are withheld: the model never sees them, and must re-find them.

2 · Train

Modelconv8·conv16·dense — yours

Loss—

3 · Scan this view

threshold0.85

Rediscovered held-out wrecks0 / 0

Rediscovery of withheld wrecks is your score — a detector that can't re-find the knowns has no business naming unknowns.

4 · File candidates

Filed candidates are hypotheses staged in candidate_targets with full provenance — model, layer, score. Verification is a survey or a dive; the simulator is where trainees fly them first.