Skip to content
anythingai.space
Reliability & Safetyintermediate13 min · 9 steps

Guardrails

Wrap a probabilistic model in input, output and action checks, know which ones actually hold, and prove they fire.

WHAT REACHES THE USERhallucinationtrust, liabilityharmful contentlegal, reputationPII leakfinesprompt injectionbreachunsafe actiondata lossbroken JSONapp crashoff-topicpoor UXUserLLM
requesterrorresponsetool calldatasignal
1 / 9
ProblemStep 1 of 9

A raw model says whatever is probable

LLMs are probabilistic. Put one straight in front of customers and some replies will be hallucinated, harmful, off-topic, malformed or leak personal data, and nothing between the model and the user notices.

Guardrails are the systems around the model that constrain what goes in, what comes out and what actions it may take.

What changed
  • +Model wired directly to users
  • +What goes wrong

Lesson 1 of 4 in Reliability & Safety