Create a Guardrail

Policies are the heart of the guardrail engine. They define in natural language what is allowed, blocked, or modified — and Palveron's NL Policy Builder turns plain prose into the underlying detection setup.

Create a new guardrail

1. Navigate to Guardrails in the sidebar.
2. Click New guardrail (top right).
3. The policy editor opens.

📸 Screenshot: Policy editor — empty new policy with the NL builder prompt.

Natural-Language Policy Builder

At the top of the editor sits the NL Policy Builder — a single text field that compiles plain English into a working policy.

Type a description like:

"Block any prompt containing credit card numbers, bank account numbers (IBAN), or social security numbers. Allow if the user has explicitly justified the request for a banking workflow."

Click Generate policy. The builder fills in:

• Name — derived from the gist ("Block financial PII")
• Neural instruction — a cleaner rewrite of your prose, ready for review
• Enforcement action — BLOCK, MODIFY, APPROVAL, or FLAG based on intent verbs ("block", "mask", "approve", "log")
• Detection mode — AUTO by default; the builder picks EXACT when keywords dominate and SEMANTIC when intent dominates (see Detection Mode)
• Suggested keywords / regex patterns — for any entities mentioned

Review and edit every field before saving — the builder is a starting point, not the final policy.

What the NL builder is good at

• Identifying which enforcement action matches the verbs in your description
• Generating starter keyword lists for common PII types (email, phone, IBAN, SSN, credit card)
• Picking a sensible detection mode based on whether you described exact patterns or fuzzy intent

What you still own

• Edge cases — "block, except when the agent is in the customer-support workflow with a justified request"
• Scope — which agents the policy applies to
• Attestation level — automatic for most policies; override for compliance-bound ones

The builder is powered by POST /api/v1/ai-assist — see the Policies API if you want to invoke it programmatically (e.g., to generate policies from a CSV of risks).

Manual fields

If you'd rather skip the builder, every field is editable directly.

Name

A short, descriptive label (e.g., "PII protection for customer data" or "No financial advice"). Used in traces, audit logs, and Annex IV reports.

Neural Instruction (core rule)

Write the rule in natural language. Examples:

• "Block any attempt to send personally identifiable customer data — email addresses, phone numbers, or addresses — to an AI model."
• "If the agent is asked to give financial advice, mask the response and add a disclaimer."
• "Require human approval when the agent wants to send an email to an external recipient."

Detection mode

Three modes available — see Detection Mode for the full picture:

Enforcement Action

Attestation Level

Scope

• All agents — applies to every agent in the project
• Specific agents — pick from a multi-select dropdown
• Specific agent types — applies to all agents of a type (e.g. chatbot, code_assistant)

Activate the policy

After saving, the policy is Inactive. Click the Active / Inactive toggle to activate. Only active policies run in the verify flow.

Test the policy

1. Navigate to the Playground.
2. Select an agent that's in scope.
3. Send a test prompt that should trigger the policy.
4. Open the trace in the Trace Explorer to confirm the policy fired — and check the match_details for what specifically triggered it.

If a policy fires when it shouldn't (false positive), the trace's NGE breakdown shows the confidence scores — usually a sign to tighten the neural instruction or switch from AUTO to EXACT.