Author Artificial Intelligence Policies
Configure Federated Guardrails permissions, inspect preview, and save an artificial intelligence policy for governance.
Use artificial intelligence (AI) authoring for Federated Guardrails protection.
The current lane configures classification scope and principal-specific Unprotect permission.
Classification scope selects the data categories that generated rules cover.
Principal scope identifies the target identities to which those rules apply.
Unprotect permits the configured inbound path to return cleartext for a principal.
Do not treat Coming Soon options as available protection methods.
Prepare access, classifications, and principals.
Confirm that your operator has configured the intended Guardrails consumer.
To author and verify an AI policy, follow these guides:
- Configure classifications and the permission matrix.
- Inspect preview and save the draft.
- Submit current saved work.
- Hand off to a separately authenticated reviewer and approver.
- Verify Guardrails publication and receipt.
Preview is not persistence.
Approval is not publication.
Consumer receipt records eligible retrieval of published output.
Consumer receipt is not evidence that every runtime has activated protection.
Here, runtime means the downstream Guardrails system evaluating traffic, not Policy’s authoring service.
Refer to states and revisions.
1 - Configure Artificial Intelligence Protection
Select classification scope and grant principal-specific Unprotect permission for Federated Guardrails.
Configure artificial intelligence (AI) protection as an authorized policy author.
Classification scope selects data categories. Principal scope selects the identities covered by generated rules.
The target principal is not necessarily the authenticated Policy author.
Before You Begin
Complete these prerequisites before configuring an AI policy:
- Complete example preparation.
- Identify one classification and one explicit supported principal.
- Confirm that your operator has configured the intended Federated Guardrails consumer.
Choose Scope and Settings
To configure the AI policy scope and settings, follow these steps:
- Open Policy Library.
- Select AI Policy from the creation choices.
Opening the form does not create a saved policy.
- Select the prepared classification.
- Select the supported principal.
- Select Federated Guardrails.
- Review the generated title.
- Enter a clearer title when needed.
Use catalog identities for users, agents, services, or groups that your deployment supports.
Do not substitute display names for native identities.
Stop if a required classification or principal cannot load.
Send the missing selection and error message to the Traverse integration administrator before submitting work.
Knowledge Graphs, Protegrity Semantic Guardrails, third-party guardrails, and Privacy
appear as Coming Soon choices.
This procedure does not enable those choices.
Set Unprotect Permission
Unprotect allows a selected principal to recover readable values from protected content.
The generated rules also protect data for the selected principals.
The Unprotect toggle controls the reverse operation. Turning it off does not disable protection.
To set permissions for each principal, follow these steps:
- Inspect the principal row in Permissions Matrix.
- Set Unprotect according to the principal’s intended permission.
Enable it to allow readable values. Otherwise leave it off.
- Review every row after changing the selected principals.
Bulk selection applies to the current principal selection.
Use it only when every selected identity needs the same Unprotect decision.
Inspect individual rows afterward.
An all-principals selection produces wildcard scope rather than a list of named identities.
Use it only with explicit authorization for that broader scope.
For a first policy, prefer an explicit principal.
Verify the Intended Matrix
Check the classification and every principal-specific Unprotect decision together.
Changing principal scope can change generated rules.
Do not infer saved or deployed permission from the matrix alone.
Next, inspect preview and save.
After Save, follow governance and
Guardrails delivery verification.
2 - Preview and Save an Artificial Intelligence Draft
Inspect Federated Guardrails preview feedback and verify that Save persists the intended draft.
Preview helps inspect current artificial intelligence (AI) authoring decisions.
It does not create a saved policy or reserve an approval token.
Preview derives output from the current inputs without publishing it to Guardrails.
Compilation checks authored inputs and derives target rules and findings.
Save runs compilation separately. It does not persist the preview response as approval evidence.
Before You Begin
Complete these prerequisites before previewing or saving:
Inspect Preview
To inspect the generated preview, follow these steps:
- Review the current classifications, principals, and Unprotect decisions.
- Inspect the preview feedback after the current inputs finish loading.
- Review any reported warnings.
- Correct scope or permissions when the feedback identifies unintended output.
Wait for feedback that corresponds to your current selections.
Do not treat a previous preview as evidence for later edits.
If preview reports an error, retain that message before changing inputs.
For a connection or directory error, send the message to the administrator of the named integration.
Save and Verify
To save the AI policy draft and verify it, follow these steps:
- Confirm a nonblank policy title.
- Select Save Draft.
Policy compiles current inputs and attempts to persist the draft.
- Inspect the Save result.
A successful Save returns you to the Library.
- Reopen the saved policy.
- Confirm its AI source, classification scope, and principal permissions.
- Record its policy identifier for review.
The saved draft is proposed work, not the governed revision containing approved intent.
Each Save records a snapshot within the proposal. The latest snapshot is its current head.
Save and preview use different operations.
A successful preview does not guarantee that Save succeeds.
A saved draft does not prove approval, publication, or consumer receipt.
If Save reports validation or compatibility errors, correct the named inputs.
If another user changed the draft, refresh current work before repeating a decision.
If the response remains uncertain, inspect the Library before creating a duplicate.
Keep the visible error and support reference for an authorized operator.
Prepare Review
Inspect saved findings before submission.
Correct hard failures rather than acknowledging them.
An authorized reviewer can inspect current saved validation through the
application programming interface (API).
That read requires review capability. Ordinary author access does not imply it.
Next, submit the current saved draft.
After approval, follow Guardrails evidence.