Model Safety

Kind: quality-attribute

Layer: Correctness Core

Record: architecture:model-safety

Severity: contextual

Scope: model-backed system, model, application

Canonical: Ontology

The degree to which a model-backed system prevents harmful inputs, outputs and actions through evaluation, guardrails and monitoring.

Listed in Architecture principles, after Explainability and before Prompt Engineering.

Requires

• Evaluation
• Guardrails
• Monitoring

Reinforces

• Model Governance
• Security

Enables

• Safe Model Deployment

Conflicts with

• Unguarded Model Autonomy

In tension with

• Capability/Utility

Tensions

• Model Safety / Capability/Utility

Violated by

• Unguarded Model Autonomy

Refactored by

• Guardrails
• Code Review
• Evaluation Suite

Severity

• contextual

Category

• Model Architecture

Reinforced by

• Artificial Intelligence Architecture
• Model Evaluation
• Agentic Architecture

Linked from

• Systems built around a model
• Correctness / Determinism / Verification
• Model Architecture
• Observability / Auditability / Traceability
• Core Vocabulary
• Model Architecture
• Quality Attributes
• Severity levels
• The resolutions