vnvspec 0.1.0

V&V-grade specifications for engineered systems — AI, optimization, simulation, and physics.

vnvspec

V&V-grade specifications for engineered systems — AI, optimization, simulation, and physics.

Overview

vnvspec lets you write machine-readable V&V specifications for any engineered system and automatically assess models against them. It produces traceable evidence reports that map to international standards (ISO/PAS 8800, ISO 21448, UL 4600, EU AI Act, and more).

Key features

• Typed spec language — Pydantic-based Spec, Requirement, IOContract, ODD, Hazard, Evidence models
• INCOSE GtWR quality checker — catch vague, ambiguous, or unverifiable requirements early
• Traceability graphs — link requirements ↔ hazards ↔ evidence ↔ standards clauses
• Standards registries — built-in clause databases for ISO/PAS 8800, ISO 21448, UL 4600, EU AI Act, NIST AI RMF
• Pluggable adapters — wrap any model (PyTorch, HuggingFace, scikit-learn, ONNX, Pyomo, FMI)
• Auto-generated tests — emit pytest and Hypothesis property tests from specs
• Rich exports — HTML reports, Markdown, GSN assurance cases (Mermaid), EU AI Act Annex IV tech docs

Installation

pip install vnvspec

With PyTorch adapter:

pip install vnvspec[torch]

Quick start

from vnvspec import Requirement, Spec

req = Requirement(
    id="REQ-001",
    statement="The classifier shall produce probabilities in [0, 1].",
    rationale="Probability outputs must be valid for downstream calibration.",
    verification_method="test",
    acceptance_criteria=["All output probabilities are between 0.0 and 1.0 inclusive."],
)

spec = Spec(name="image-classifier-v1", requirements=[req])
violations = req.check_quality()
print(f"Quality issues: {len(violations)}")

Development

# Install all dev dependencies
just install

# Run all checks (lint, typecheck, test with coverage)
just check

License

Apache-2.0. See LICENSE.