Scanner & Governance

The Scanner & Governance page controls how CleverGuard inspects every message flowing through your agents — what it detects, how hard it looks, and what happens when something is found. Changes apply immediately to every agent in your organization.

Key Concepts

The CleverGuard pipeline

Every message sent to or received from an agent passes through a four-tier scanning pipeline before it is forwarded. Each tier runs in sequence; a high-confidence finding at an early tier short-circuits the later ones so you only pay latency for what is needed.

Tier 0Data labels

Deterministic label rank check. Enforces org-defined sensitivity classifications (Purview, Google DLP, Macie, custom). Runs first, adds under 1ms, zero false positives.

Tier 1Pattern match

Regex and rule-based detection for PII (SSNs, emails, phone numbers), API keys, PATs, connection strings, and common injection signatures. Fast and auditable.

Tier 2CleverGuard + ScopeGuard

Clevername's fine-tuned Gemma LoRA models. CleverGuard catches injection and jailbreak attempts; ScopeGuard catches scope violations — agents doing things outside their approved mandate. Runs in ~30ms on cache hit.

Tier 3Contextual LLM (optional)

Hand-off to Clevername's self-hosted judge for governed scans and ambiguity/disagreement escalation. It does not use your provider key and has no managed-provider fallback.

What the settings page controls

Tier 3 escalation — enable the optional self-hosted contextual judge when available. There is no provider picker or customer Tier 3 key path.
Content safety rules — toggle individual scanner categories (PII, injection, secrets, toxicity, URL blocking) and set each to standard or strict mode.
Detection response — choose whether a detection triggers Visibility (log only), Notify (log + email admins), or Enforce (block the request or restrict the agent).
ClaimGuard output verification — verify that agent responses are grounded in the source context. Flags or blocks unsupported or contradicted claims before they reach users.

Note

Settings here are org-wide defaults. Individual agents can override content safety settings through their guardrail profile (compiled from the Agent Review questionnaire). The org default applies to any agent that does not have an override.

Reviewing Tier 3 escalation

Open Scanner & Governance

Navigate to Settings → Scanner & Governance. The Tier 3 card explains the current self-hosted beta governance-compute behavior near the top of the page.

Check availability

Tier 3 is operated by Clevername, not by a customer BYOK provider. If the backend endpoint is unavailable, scan results surface reduced coverage rather than silently pretending the contextual judge ran.

Confirm key separation

Your normal agent and chat calls still use your configured provider keys. Tier 3 does not read those keys, does not spend them, and does not fall back to a managed provider if the self-hosted judge is unavailable.

Use response modes for enforcement

Tier 3 contributes findings to the same Visibility, Notify, and Enforce response modes as the rest of CleverGuard. Use Enforce only after you have reviewed alert volume.

Monitor reduced coverage

If Tier 3 cannot run, downstream surfaces show reduced coverage signals such as ml_degraded. Treat those as an operational signal, not a clean Tier 3 verdict.

Detection response modes

Choose how CleverGuard responds when it detects a threat across your org:

VisibilityLog the detection as a DLP event in the SOC console. Nothing is alerted and nothing is blocked. Use this when you are calibrating or want full observability without friction.

NotifyLog the detection and send an email alert to org admins. The agent keeps running. Good for teams that want to be in the loop but do not yet need automated enforcement.

EnforceBlock the offending request or, for drift events, restrict the agent. For drift specifically, enforcement only activates once the agent is governed — either via human-approved questionnaire or after reaching the observation threshold you set.

CleverGuard deep dive Detection response & enforcement ClaimGuard output verification Agent Review & guardrail profiles