Certificate-Based Auditing for Reproducible Drift Detection: An Empirical Study in Time-Series Forecasting
ghost-drift-audit-jp is an audit engine that fixes drift detection (distribution shift / regime shift) in operational time-series forecasting as a reproducible protocol. It outputs the split boundaries, threshold policies, input data identification, execution code, and runtime environment as a unified certificate, so that a third party can regenerate the same audit verdict (OK/NG) from the same inputs. In particular, estimation is restricted to the Calibration phase and the Test phase is used only for evaluation, structurally eliminating post-hoc threshold tuning (after-the-fact optimization) once results are observed. As a case study, we target electric power demand × weather time-series data (Jan–Apr 2024), generate certificates, ledgers, and evidence time series, and present the audit verdict as reproducible artifacts.
To address the “opaque inference” problem in conventional AI operations, this framework provides the following.
[!TIP] Audit-First Design Alongside running predictions, it automatically generates objectively verifiable evidence for third parties.
[!IMPORTANT] Tamper-evident Fingerprints It fixes hash fingerprints of input data and configuration parameters, making post-hoc modifications mathematically detectable.
[!NOTE] Verifiable Integrity Rather than mere statistical optimality, it makes visible the model’s faithful adherence to operational rules.
.
├── ghost_drift_audit_JP.py # Core Logic & Audit Engine
├── electric_load_weather.csv # Input: Weather (Synthetic)
├── power_usage.csv # Input: Demand (Synthetic)
└── adic_out/ # Output: Accountability Ledger
Switch the strictness of the audit via AUDIT_CONFIG['PROFILE'].
| Profile | Use / Target | Strictness | Key Features |
|---|---|---|---|
| demo | Smoke test / learning | Low | Prioritizes understanding behavior and evidence output |
| paper | Research / reproducible experiments | Mid | Ensures computational reproducibility via fixed seeds |
| commercial | Production / decision-making | High | Produces strict gate checks and a final verdict |
AUDIT_CONFIG = {
"PROFILE": "demo", # "demo" | "paper" | "commercial"
}
pip install numpy pandas matplotlib lightgbm
Place the CSV files in the same directory as the .py.
[!CAUTION] The bundled CSVs are synthetic (dummy) data. They are for smoke testing only; for production use or research, use real data for which you hold the rights.
python ghost_drift_audit_JP.py
From “prediction” to “accountability.” Produced by GhostDrift Mathematical Institute (GMI) — Official Website | Online Documentation