> ## Documentation Index
> Fetch the complete documentation index at: https://evalgate.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent governance

# Enterprise governance and compliance controls

> Configure provider policies, PII protection, retention, and access controls for your org from the System Controls card in Settings → Workspace.

EvalGate enforces governance as system behavior, not as a checklist. Provider restrictions, PII scrubbing, drift detection, and audit logging are part of the runtime path — they run on every evaluation and judge execution, not just when you remember to check them. Configure the editable controls from the **System Controls** card in **Settings → Workspace** (`/settings/organization`).

> [accordion] Provider and model policy
>
> Control which AI providers and models your organization can use for judge evaluations. These settings apply to all evaluation runs across your org.
>
> **What you can configure:**
>
> * **Allow external LLM providers** — one switch that blocks or allows outbound model calls at the org boundary
>
> * **Provider allowlist** — restrict outbound calls to an approved list of providers; an empty list allows none
>
> * **Cost alert budgets** — daily, weekly, and monthly USD budgets (`costAlerts` on `PUT /api/system/controls`)
>
> > [!NOTE]
> >
> > Provider and model restrictions are enforced before outbound calls are made. Disallowed providers are blocked at the platform boundary — the request never reaches the external API.

> [accordion] PII protection
>
> EvalGate scrubs inputs before they reach the embedding pipeline or external judge calls, so sensitive data in your test cases or traces does not leave your org boundary in cleartext.
>
> **How it works:**
>
> * Inputs pass through the scrubber before embedding or being sent to an external judge
> * External judge calls respect your org's PII policy — if scrubbing is not permitted for a given request, the provider call is blocked entirely
> * Scrubbing is auditable: the audit log records that scrubbing occurred and what policy applied, without storing the raw sensitive values themselves
>
> **To enable PII protection:**
>
> In **Settings → Workspace → System Controls → PII policy**, choose a scrub mode (Redact, Tokenize, Hash, or Drop) and decide whether to allow PII in embeddings and raw PII storage. Changes apply to subsequent runs.
>
> > [!WARNING]
> >
> > Enabling PII protection does not retroactively scrub data already stored in traces or evaluation records. Apply PII policies before ingesting sensitive data.

> [accordion] Drift enforcement
>
> The drift gate enforces cross-surface parity — it detects when evaluation behavior shifts across environments, model versions, or prompt changes and treats those shifts as violations, not informational alerts.
>
> **What drift enforcement covers:**
>
> * **Cross-surface parity** — the drift gate compares results across surfaces (e.g. staging vs. production, model v1 vs. v2) and flags deviations that exceed your thresholds
>
> * **Runtime monitors** — continuous tracking of score distribution and variance shifts catches gradual degradation that single-run comparisons miss
>
> * **Hardcoded threshold violations** — if you set a minimum pass rate or quality score floor, breaches are treated as violations and surface in the dashboard and CI gates
>
> * **Versioned, auditable baselines** — every baseline is versioned so you can trace exactly which run and configuration a comparison was made against
>
> > [!TIP]
> >
> > Use drift enforcement in combination with CI gates to automatically block deployments when a model or prompt change causes a statistically significant quality regression.

> [accordion] Audit logging and org isolation
>
> Every action that modifies data in EvalGate is org-scoped and recorded in the audit log. This applies to all users and API keys in your organization.
>
> **What is audited:**
>
> * All mutations (evaluation creation, run creation, test case changes, configuration updates) are attributed to an org member or API key with a timestamp
> * Judge executions record the full execution context: model config, retry count, timing, and whether PII redaction was applied
> * Controls are queryable through the API so you can export audit data for compliance reporting
>
> **Access control:**
>
> * RBAC is part of the control layer — role assignments determine which members can create evaluations, run judges, export data, and modify org settings
> * Multi-tenant isolation ensures data from one org is never accessible to another, even under the same EvalGate instance
>
> Manage member roles from **Settings → Team**.

> [accordion] Retention and destructive cleanup
>
> Retention policy in EvalGate is not just metadata tagging. You can configure org-scoped destructive retention sweeps that permanently remove expired evaluation data and logs under policy.
>
> **How retention works:**
>
> * Set a retention period for evaluation runs, traces, and audit records
> * When a sweep runs, expired data is actually deleted — not soft-deleted or flagged
> * Each sweep is recorded in the audit log with scope, timestamp, and affected record counts so you have full visibility into what was removed and when
>
> Configure eval data retention, the audit evidence window, Copilot message retention, and remote runner log retention in **Settings → Workspace → System Controls**.
>
> > [!WARNING]
> >
> > Destructive retention sweeps are irreversible. Review your retention settings carefully and confirm the policy applies only to the data categories you intend before activating a sweep.

## Where to configure governance

Governance controls live under **Settings**. Updating System Controls requires an Admin role or higher; drift thresholds come from the org policy and are shown as drift alerts, not edited here.

> [card] Provider and PII policy
>
> Settings → Workspace → System Controls

> [card] Runtime policy packs
>
> Settings → Rules

> [card] Activity history
>
> Settings → Activity history

> [card] Member roles
>
> Settings → Team


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.