Glossary: Guardrail
Guardrail, on this site, means a decision contract that screens untrusted text (prompt, completion, or tool call) with Jev Noul / Score / Choice questions, then lets application code allow, review, or block. It is not a certified security product, not a TypeSafe firewall SKU, and not a promise that schema-safe answers are factually correct. Jev is TypeSafe’s System One decision model (POST /v1/systemone).
This glossary entry is for builders who already heard of Jev / System One and need a citable definition before they replace allowlists with a single Noul. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. Confirm the LLM guardrails cookbook on docs.typesafe.ai. We do not sell or issue API keys.
How it shows up
It is not TypeSafe’s product name for a firewall. Official docs describe the pattern under the LLM guardrails cookbook. You still own the verdict: your service returns 403, holds the tool, or queues review. jev-1.13 is not hardened against adversarial state by default — TypeSafe’s jaggedness note calls that out.
Typical screen: a Noul “is this a policy violation?” plus a threshold on noul (Noul has no confidence field). High-stakes blocks may add a Choice among violation classes. Pin jev-1.13.0 once floors are fitted; jev-latest can move. Log the full answers in a decision trace. Hard constraints (account allowlists, amount caps) stay deterministic. Language-y gray zones become questions.
Schema validation (422 on a bad body; closed criteria on the answer) does not prove the screen was right. A valid Noul can still miss a novel jailbreak.
Contrast
Guardrail vs decision contract. The contract is the triple you ship: state shape, questions, and code floors. A guardrail is one use of that contract on untrusted text.
Guardrail vs schema validation. Schema validation keeps the HTTP body and answer keys legal. A guardrail is a product policy on content.
Guardrail vs abstention. Abstention is “do not auto-act.” A guardrail often fail-closes (block) when unsure; that is a policy choice, not a Jev default.
Guardrail vs LLM-as-judge essays. Jev returns typed slots. It does not write a safety report. If you generate a paragraph for auditors, label it assistance.
Common mistakes
- Expecting Jev to block the HTTP request for you.
- Replacing allowlists and amount caps with one Noul.
- Assuming
jev-1.13is adversarially hardened. - Copying a Choice confidence floor onto a Noul screen.
- Inventing
/v1/moderateas an official route. Moderate is a pattern you implement.
See also
FAQ
Does Jev block the request for me?
No. You receive probabilities. Your service returns 403 or holds the tool.
Is this a replacement for allowlists?
No. Hard constraints stay deterministic. Language-y gray zones become questions.
Is guardrail a TypeSafe product name?
Not in the public API we read. Official docs discuss an LLM guardrails cookbook. This page is a working definition.
Should I fail open when Jev is down?
Not on irreversible or safety-critical acts. See fallback.
Limits
This is a jev.pro working definition unless TypeSafe uses the same word as an API field. We do not invent search volume. Schema-safe answers are not automatically factually correct. This page is not a security certification.
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.