Grade Prediction Calibration Index (GPCI)

GMI-MIN-005MiningIntegration-Required

Definition

Degree to which AI orebody model grade predictions correspond to actual reconciled grades, measured by comparing predicted grades to blast hole assay and mill head grades.

The Measurement Gap

Traditional grade control reconciliation compares predicted grades to actual grades using aggregate metrics: mean error, mean absolute error, and correlation coefficients across the entire prediction set. These metrics tell the geology team whether the model is biased (systematic over- or under-prediction) and how large the typical error is. What they cannot answer is whether the model knows when it is uncertain.

A model that reports 90% confidence on every prediction and achieves 70% accuracy everywhere is poorly calibrated but will appear reasonable under aggregate error metrics. The geology team making decisions in the pit needs to know whether a "high confidence" prediction is meaningfully more reliable than a "low confidence" one. Legacy metrics like ore-to-waste reconciliation ratios, MTBF of mining equipment, PM Compliance, and cost per ton are entirely blind to model calibration. Even standard geostatistical validation metrics (cross-validation error, variogram reproduction) test model structure, not the reliability of the confidence scores that drive operational decisions.

Decision Supported:

Can the geology team trust the model's confidence scores when making grade control decisions, or is the model systematically overconfident or underconfident?

Formula

Plain Language: Group all grade predictions into confidence bands (for example, 60-70%, 70-80%, 80-90%, 90-100%). Within each band, measure how often the prediction was actually correct (within a site-specific tolerance). Divide that actual accuracy by the stated confidence. If the model says "I am 90% confident" and it is right 90% of the time, the ratio is 1.0 (perfectly calibrated). If it is right only 65% of the time, the ratio is about 0.72 (overconfident). Compute a weighted average across all bands for the overall score.

Notation:

GPCI(b) = Actual_accuracy(b) / Stated_confidence(b)

GPCI_weighted = Sum(n_b x GPCI(b)) / Sum(n_b)

Variable Definition Unit
b Confidence band (e.g., [0.6-0.7), [0.7-0.8), [0.8-0.9), [0.9-1.0]) Dimensionless range
Actual_accuracy(b) Proportion of predictions in band b where the absolute error between predicted and actual grade is within tolerance Ratio (0 to 1)
Stated_confidence(b) Midpoint of confidence band b Ratio (0 to 1)
n_b Number of predictions in band b Count
Tolerance Site-specific acceptable prediction error (e.g., +/-10% relative or +/-0.05% Cu) Grade units

A result of 1.0 means perfectly calibrated. Above 1.0 means underconfident (model is better than it claims). Below 1.0 means overconfident (dangerous for operational decisions).

Proposed formula, pending validation pilots.

At a Glance

Unit
Ratio
Trend Direction
Toward 1.0
Reporting Frequency
Monthly
Category
Predictive Geology Model Integrity
← Geological Domain Drift Score (GDDS)Model Boundary Attribution Rate (MBAR) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment