Live scan: see the ore, then decide
A real hyperspectral scan builds up line by line. Absorption maps show what the camera sees; the model predicts with an honest interval; the policy decides.
Scan queue
Pre-registered showcase samples (chosen by hash, not by result)
—
—
Audit log
Every outcome, with model, calibration and policy hashes
Before grinding
Deployable model · split-conformal 80% interval
Why
Wavelength regions the model leaned on (occlusion, pipeline recomputed): association, not mechanism
Belt XRF reading
Chromitite parcel
Downhole log
Logged seam, Cr₂O₃ and 4E grade against the out-of-fold prediction (shaded: 80% interval)
What the evidence says
Grouped by project (123 projects, 1,112 intervals); three gates
From chemistry to PGE
Out-of-fold · 80% interval · project-held-out
Silica in concentrate: forecast vs lab vs last assay
Real plant, September 2017 test month, hourly
Real plant parameters, this hour
Hourly means of 20 s tags (Kaggle, CC0)
Decision for the metallurgist
Decision support only. Relationships are correlational; a controlled trial comes before any setpoint moves.
Honest result
Test month, after leakage controls
Import lab results (QEMSCAN · XRF · XRD · assays)
Typed against the versioned target registry; incompatible rows are rejected with a reason
Reconciliation: prediction vs lab
Retrospective: the lab values are the public labels the models were scored against
Export to plant and lab systems
Files, not live connections. Each one says what it is.
Predictions, intervals and decisions per sample. Text that could start a spreadsheet formula is neutralised.
Model, calibration and data hashes, policy version and evidence bar.
NodeId, value, quality, timestamp, for a site historian or OPC UA server to read. This is not a live server.
Parcel records with null geometry and a source_location_id left for the fleet-management system: HIDSAG has no coordinates and none are invented.
Printable, with every number placed by code.
Target registry
—