Dataset and Replication Package for M-11: Auditing Synthetic Risk Labels and Physical Baselines in a Cyber-Physical Benchmark for Underground Mining

Overview: This deposit hosts the official replication package, synthetic benchmark dataset, and experimental audit artifacts for the scientific manuscript: "M-11: Auditing Synthetic Risk Labels and Physical Baselines in a Cyber-Physical Benchmark for Underground Mining" (submitted to Acta Montanistica Slovaca, 2026). M-11 is a simulation-only cyber-physical benchmark that integrates AI4I 2020 predictive maintenance attributes with synthetic kinematic positioning, physiological, and environmental indicators (10,000 records; 37 columns) to audit the statistical mechanics and potential circularity of synthetic risk labelling in confined underground mining operations. Key Empirical Findings Reported in the Manuscript: Label Circularity Audit: Variance decomposition demonstrates that 82.07% of the synthetic score variance is driven by spatial/kinematic components, 17.43% by injected noise, 0.48% by multimodal variables, and 0.02% by cross-covariance. Symmetric Holdout Evaluation (n = 2,000; seed 42): Calibrated Time-to-Collision (TTC) and Random Forest both achieve F2 = 0.9103; XGBoost achieves F2 = 0.9065. The paired bootstrap difference (ΔF2 = -0.0038, 95% CI: -0.0129 to 0.0047) and Holm-adjusted McNemar test (p = 0.6015) confirm no statistically significant gap under circular labels. Decoupled Hazard Stress Tests: Under 22% class prevalence and low signal-to-noise ratio, non-spatial decoupled labels (CO, fatigue, vibration) push F2 optimisation to near-universal HIGH alerts, approaching the trivial Always-HIGH baseline (F2 = 0.5851). A multimodal-only composite target recovers F2 = 0.6408 with 54.1% specificity. Feature Ablations: Removing three direct kinematic indicators reduces F2 by 9.60 percentage points; removing the entire 8-feature spatial block reduces F2 by 32.52 percentage points from the full 0.9065 baseline. Computational Scalability: Local backend P95 latency scales to 27.91 ms under 500 concurrent simulated agents (< 200 ms supervisory threshold). The WebGL/Three.js integrated-GPU prototype achieves 34.8 FPS under nominal load. Deterministic Safety & GenAI Audit: The independent deterministic safety layer satisfies 17 automated unit/integration tests and 100/100 scripted regulatory scenario checks. The Gemini language model acts strictly as an exploratory natural-language interface without numerical autonomy. Package Contents & Cryptographic Verification (SHA-256): public_mining_equipment_dataset.csv (3,859,401 bytes) — SHA-256: b71d0b6baed8b6e0bc7eee1a1a5b7440d06a04ddaea5d502907cf03d958705f0 requirements.txt (574 bytes) — Frozen Python 3.10.11 scientific environment specifications. results/feature_dictionary.csv (2,928 bytes) — Full data dictionary defining all 37 attributes, physical units, and provenance. results/label_circularity_audit.json (17,530 bytes) — Machine-readable variance decomposition and baseline comparison exports. results/genai_audit_bank.json (35,882 bytes) — 100 structured synthetic scenarios for LLM compliance auditing. results/genai_audit_review_sheet.csv (47,295 bytes) — Manual verification rubric with inter-annotator agreement (Cohen's κ = 0.85). Boundary of Scope: M-11 is a computational simulation benchmark and exploratory supervisory prototype. It does not establish mine-site functional safety, nor does it claim active-mine physical twin synchronisation or autonomous braking actuation.