08 · Safety & Alignment

Guardrails

Safety constraints around a model

The safety constraints and filters layered around a model to block harmful inputs or outputs. Often a separate system sitting in front, distinct from the model's own training.

Concrete example

A filter that catches a dangerous request before it ever reaches the model, or scrubs an unsafe reply after.

Why it matters

Becomes critical the moment an agent can take real actions, not just talk.