Model Inference

Kind: capability

Layer: Correctness Core

Aliases: Runtime Prediction/Generation

Record: architecture:model-inference

Severity: contextual

Scope: service, model serving

Canonical: Ontology

The ability to run a trained model on validated input at runtime and return a checked prediction or generation.

Listed in Architecture principles, after Model Evaluation and before Retrieval-Augmented Generation (RAG).

Requires

• Model Artifact
• Input/Output Contract

Reinforces

• Artificial Intelligence Architecture

Conflicts with

• Training-Time-Only Model Logic

In tension with

• Latency/Cost

Tensions

• Model Inference / Latency/Cost

Violated by

• Opaque Ungoverned Model Use

Refactored by

• Extract Module
• Extract Adapter

Severity

• contextual

Category

• Model Architecture

Required by

• Prompt Engineering
• Agentic Architecture

Linked from

• Model Architecture
• Codebase / System Architecture Styles
• Core Modular Design
• Model Architecture
• Severity levels
• The resolutions