🎓 Lesson 15
D5
Audit Trails for AI-Driven Resource Statements
An audit trail for AI-driven resource statements is a complete, chronological record of every step, decision, data input, and model change used to generate a mineral resource estimate—so anyone can verify how and why the result was produced.
🎯 Learning Objectives
- ✓ Explain the regulatory and ethical necessity of audit trails in AI-generated resource statements
- ✓ Design a minimal viable audit trail architecture for a grade-control AI workflow
- ✓ Analyze an AI resource report to identify missing or non-compliant audit trail elements
- ✓ Apply version control and metadata tagging standards to track model iterations and data provenance
- ✓ Evaluate the completeness and integrity of an audit trail using the CRIRSCO-aligned validation checklist
📖 Why This Matters
In 2023, a major Australian iron ore operator revised 12.4 Mt of Measured Resources after discovering their AI grade estimation model had silently retrained on outdated drill-hole assays—without logging the data refresh. No audit trail meant no way to reconstruct when, why, or how the change occurred. For mining engineers, audit trails aren’t bureaucratic overhead—they’re the bedrock of trust: they protect license to operate, uphold investor confidence, and ensure technical defensibility during audits, litigation, or public disclosure. Without them, AI doesn’t augment judgment—it obscures it.
📘 Core Principles
Audit trails rest on three interlocking pillars: provenance (who provided what data, when, and in what form), process fidelity (exact software versions, parameters, and preprocessing logic applied), and decision lineage (human reviews, classification rationale, and uncertainty flags). In AI-driven resource workflows, this extends beyond traditional geostatistics to include model cards (documenting training data bias, performance metrics per lithology), SHAP values for local interpretability, and containerized execution environments (e.g., Docker images with pinned library versions). Regulatory frameworks like JORC (2012) Clause 22.2 and NI 43-101 Form 43-101F1 explicitly require ‘sufficient documentation to support the reasonableness’ of estimates—making audit trails not optional, but legally material.
📐 Audit Trail Completeness Index (ATCI)
The ATCI quantifies how fully an AI resource statement’s audit trail satisfies CRIRSCO-aligned minimum requirements. It scores traceability across five dimensions—data provenance, model versioning, parameter transparency, human review logging, and uncertainty propagation—with each scored 0–1 and averaged. A score < 0.7 triggers mandatory remediation before sign-off.
Audit Trail Completeness Index (ATCI)
ATCI = (P + V + T + R + U) / 5Quantitative measure of audit trail adherence across five essential dimensions, expressed as a normalized score between 0 and 1.
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| P | Data Provenance Score | unitless | 0–1 score reflecting completeness of source data metadata, versioning, and QA/QC traceability |
| V | Model Versioning Score | unitless | 0–1 score reflecting immutability, reproducibility, and environment capture of AI models |
| T | Parameter Transparency Score | unitless | 0–1 score reflecting full disclosure of all tunable parameters, preprocessing steps, and algorithmic assumptions |
| R | Review & Approval Log Score | unitless | 0–1 score reflecting timeliness, authorship, and technical justification of human interventions |
| U | Uncertainty Propagation Score | unitless | 0–1 score reflecting explicit linkage between AI prediction uncertainty and resource classification confidence |
Typical Ranges:
Regulatory minimum compliance: 0.70 – 0.79
Industry best practice: 0.90 – 0.98
💡 Worked Example
Problem: A copper porphyry resource model uses AI kriging. Audit logs show: (1) Drill data sourced from 2021–2023 database (versioned, timestamped); (2) Model v2.3.1 (Docker hash: a1b2c3); (3) All 12 hyperparameters logged, including variogram range = 85 m; (4) Two geologist reviews logged with timestamps and comments; (5) Local kriging variance mapped but no conditional simulation uncertainty reported. Score each dimension: 1=1.0, 2=1.0, 3=1.0, 4=1.0, 5=0.4.
1.
Step 1: Assign binary scores per dimension (1 = compliant, 0 = missing, 0.5 = partial). Here: [1.0, 1.0, 1.0, 1.0, 0.4].
2.
Step 2: Compute arithmetic mean: (1.0 + 1.0 + 1.0 + 1.0 + 0.4) / 5 = 4.4 / 5 = 0.88.
3.
Step 3: Compare to threshold: 0.88 > 0.7 → audit trail is compliant but requires uncertainty enhancement for best practice.
Answer:
The ATCI is 0.88, which exceeds the minimum compliance threshold of 0.7 but falls short of the recommended best-practice target of ≥0.95.
🏗️ Real-World Application
At Newmont’s Boddington Gold Mine (Western Australia), AI-grade control models now run within a validated pipeline where every prediction triggers an auto-generated audit bundle: (a) SHA-256 hash of input assay composite file, (b) JSON manifest listing scikit-learn version, random seed, and feature importance ranking, (c) PDF report signed by the Competent Person showing kriging variance vs. AI prediction error bands, and (d) blockchain-stamped timestamp (using Hyperledger Fabric) immutably anchored to the corporate data lake. This system reduced regulatory query resolution time from 17 days to <2 hours during the 2022 JORC audit.
📋 Case Connection
📋 Copper Mine Block Model Refinement Using Neural Kriging
Traditional kriging over-smoothed high-grade chalcocite zones, causing 8.2% reserve underestimation