dev404ai/governance-benchmark-dataset

Governance Benchmark Dataset

Curated benchmark dataset and evaluation harness for comparing governance evidence feasibility across four decision system architectures: deterministic rule engines, hybrid ML+rules, streaming feature-driven systems, and agentic AI.

This dataset is the artifact mirror of the benchmark described in "Governed Auditable Decisioning Under Uncertainty: Synthesis and Agentic Extension" (arXiv:2604.19112). The canonical code repository is on GitHub and the citable archival deposit is on Zenodo (links below).

What it contains

Eight curated decision scenarios (two per architecture), each carrying a complete decision-event-schema record (schema_version 0.3.x), a per-property feasibility matrix, and — for the hybrid and agentic architectures — cascade traces of multi-step decision chains.

Path	Content
dataset/architectures/<arch>/scenarios/*.json	Scenario records with embedded decision events
dataset/architectures/<arch>/metadata.json	Architecture description and collapse modalities
dataset/cascade_traces/*.json	Multi-step decision chains (hybrid, agentic)
dataset/feasibility_matrices/cross_architecture.parquet	48-row per-property feasibility table
dataset/schemas/*.json	JSON Schemas for scenarios, traces, and matrices
dataset/CHECKSUMS.sha256	sha256 pins for every committed data file

The feasibility matrix scores each decision-event property per scenario (architecture_type, scenario_id, property_name, level, recovery_cost, fillable_ratio, opaque_ratio).

Loading

The flat feasibility matrix loads directly with datasets:

from datasets import load_dataset

matrix = load_dataset("dev404ai/governance-benchmark-dataset", "feasibility_matrix")

Scenario records are nested documents; load them with the typed loaders from the canonical repository (validation, frozen dataclasses, scoring included):

# pip install "git+https://github.com/governance-evidence/governance-benchmark-dataset"
from pathlib import Path
from benchmark.loaders.scenario import load_scenarios

scenarios = load_scenarios(Path("dataset/architectures/agentic_ai/scenarios"))

Integrity

Every data file is pinned in dataset/CHECKSUMS.sha256; the canonical repository enforces the pins in CI. Verify a download from the mirror root with:

shasum -a 256 -c dataset/CHECKSUMS.sha256

Scope and responsible use

Scenarios are curated, synthetic-but-realistic constructions designed to expose architecture-specific structural breaks (decision diffusion, evidence fragmentation, responsibility ambiguity); they are not records of real incidents or real persons. The benchmark measures the feasibility of governance evidence, not the quality of any production system, and is not a compliance certification instrument.

Licensing

• Dataset files: CC BY 4.0
• Code in the canonical repository: Apache-2.0

Citation

@misc{solozobov2026governedauditable,
  author = {Solozobov, Oleg},
  title  = {Governed Auditable Decisioning Under Uncertainty: Synthesis and Agentic Extension},
  year   = {2026},
  eprint = {2604.19112},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CY},
  doi    = {10.48550/arXiv.2604.19112},
  url    = {https://arxiv.org/abs/2604.19112}
}

Software/dataset deposit (concept DOI, resolves to the latest release): 10.5281/zenodo.19248722.

Related projects

Repository	Role
decision-event-schema	Schema whose properties this benchmark scores
evidence-sufficiency-calc	Sufficiency scoring used in benchmark evaluation
governance-drift-toolkit	Drift monitoring validated by benchmark scenarios
evidence-collector-sdk	Produces decision events conforming to the schema

Provenance

This mirror is synced from tagged releases of governance-evidence/governance-benchmark-dataset by a sync script; the GitHub repository is the source of truth.