Governance Drift Toolkit

A Python toolkit for label-free monitoring of governance evidence degradation in risk decision systems. Answers: "Is our governance evidence still sufficient, even though we can't see the ground truth yet?"

The toolkit combines proxy drift monitors, composite alerting, harmful-shift suppression, and a governance response chain for delayed-label environments such as fraud detection, credit scoring, and related risk systems.

Academic Context

This toolkit is shared across two papers in the governance-evidence series.

Primary (label-free extension, current focus):

Solozobov, O. (2026). Label-Free Detection of Governance Evidence Degradation in Risk Decision Systems. arXiv:2604.17836. https://arxiv.org/abs/2604.17836

Companion (introduces the drift toolkit as part of the sufficiency framework):

Solozobov, O. (2026). Evidence Sufficiency Under Delayed Ground Truth: Proxy Monitoring for Risk Decision Systems. arXiv:2604.15740. https://arxiv.org/abs/2604.15740

Synthesis context — this toolkit is one of the artifacts whose transferability across decision system architectures is assessed in:

Solozobov, O. (2026). Governed Auditable Decisioning Under Uncertainty: Synthesis and Agentic Extension. arXiv:2604.19112. https://arxiv.org/abs/2604.19112

Install

From a Package Index

Use this when the package is published to your package index:

pip install governance-drift-toolkit

From GitHub

Use this before package-index publication, or when installing directly from source control:

pip install git+https://github.com/governance-evidence/governance-drift-toolkit.git

Optional sufficiency integration:

pip install "governance-drift-toolkit[sufficiency]"

For Contributors

Clone the repository, create a local virtual environment, and install development dependencies:

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

For dataset-backed demos, install the extra demo dependencies:

pip install -e ".[dev,demo]"

Dataset download and local directory setup for the demos are documented in docs/demo_datasets.md.

Quick Start

The core package depends only on NumPy and SciPy. A minimal monitoring pass looks like this:

import numpy as np
from drift import (
    compute_composite_alert,
    determine_response,
    fraud_detection_config,
)
from drift.monitors.score_distribution import compute_psi
from drift.monitors.feature_drift import compute_feature_psi
from drift.monitors.uncertainty import compute_prediction_entropy

config = fraud_detection_config()
rng = np.random.default_rng(42)

ref_scores = rng.normal(0.30, 0.15, size=1000)
cur_scores = rng.normal(0.45, 0.20, size=1000)

results = [
    compute_psi(ref_scores, cur_scores),
    compute_feature_psi(
        rng.normal(size=(500, 3)),
        rng.normal(0.5, 1.0, size=(500, 3)),
    ),
    compute_prediction_entropy(rng.uniform(0.1, 0.9, size=500)),
]

alert = compute_composite_alert(results, config)
response = determine_response(alert, config)
print(f"Alert: {alert.severity.value}, Response: {response.action.value}")

See docs/deployment.md for installation modes and docs/alerting.md for the composite alert logic.

For dataset-backed demos, see docs/demo_datasets.md.

Seven Proxy Monitors

#	Category	Detects	Misses
1	Score Distribution Shift	P(X) changes in scores	Adversarial drift preserving scores
2	Feature Drift	Covariate shift in inputs	Concept drift with stable features
3	Uncertainty	Calibration degradation	Confident-but-wrong predictions
4	Cross-Model Disagreement	Adversarial evasion	Correlated model failures
5	Operational Process	Behavioral changes	Fast-onset drift
6	Outcome-Maturity	Cohort-based drift	Novel patterns
7	Proxy Ground Truth	Pattern changes pre-labels	Social engineering

Governance Response Chain

Monitor -> Alert -> Escalate -> Fallback -> Rollback

Related Projects

This toolkit is part of the governance-evidence toolkit:

Repository	Role	Concept DOI
decision-event-schema	Schema for events this toolkit monitors	10.5281/zenodo.18923177
evidence-sufficiency-calc	Sufficiency scoring — bidirectional integration with this toolkit	10.5281/zenodo.19233930
evidence-collector-sdk	Collects evidence streams that feed into this toolkit	10.5281/zenodo.19245404
governance-benchmark-dataset	Cross-architecture benchmark that validates this toolkit's scenarios	10.5281/zenodo.19248722

All DOIs above are concept DOIs -- each resolves to the latest Zenodo release of that artifact.

Citation

If you use this toolkit in your research, please cite both the paper and the software artifact.

Paper (primary):

@misc{solozobov2026labelfree,
  author = {Solozobov, Oleg},
  title  = {Label-Free Detection of Governance Evidence Degradation in Risk Decision Systems},
  year   = {2026},
  eprint = {2604.17836},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CY},
  doi    = {10.48550/arXiv.2604.17836},
  url    = {https://arxiv.org/abs/2604.17836}
}

Software (this repository):

@software{solozobov2026governancedrift,
  author  = {Solozobov, Oleg},
  title   = {Governance Drift Toolkit},
  version = {0.2.1},
  year    = {2026},
  url     = {https://github.com/governance-evidence/governance-drift-toolkit},
  doi     = {10.5281/zenodo.19236417}
}

The software doi above is the concept DOI (always resolves to the latest Zenodo release). The current v0.2.1 version DOI is 10.5281/zenodo.19248601.

See CITATION.cff for machine-readable citation metadata.

License

Apache-2.0