Real-time policy engine
Policies evaluate in-line, before the response reaches your user — built for conversational latency budgets.
GD
Runtime guardrails
BetaStop bad outputs before they ship.
NovaGuard puts your evaluation criteria in front of live traffic: block unsafe responses, enforce policy, and escalate to humans in real time — using the same scorers you already trust offline. Now in private beta.
Real-time
scoring on live traffic
1:1
parity with offline scorers
Beta
design-partner program
NovaGuard · policy engineLIVE
pii_disclosureBLOCK
refund_above_limitESCALATE
hallucination_risk > 0.2ESCALATE
brand_tone_checkALLOW
4 policies enforced on live traffic · every decision audit-logged
01 — Capabilities
Your eval criteria, enforced in production.
Block unsafe output
PII disclosure, policy violations, unsafe content: stopped at the gate with a deterministic action — a configurable fallback response or a rewrite-and-retry.
Same scorers as your evals
No second quality standard. The criteria you evaluate with offline are the criteria NovaGuard enforces in production.
Human escalation
Route uncertain or high-stakes cases to a human queue with full trace context attached.
Compliance rule packs
Prebuilt policy packs for regulated conversations — financial disclosures, healthcare boundaries, consent language — drop in rather than author from scratch.
Full audit trail
Every allow, block, and escalation is logged with its score and reasoning — evidence for auditors, not just dashboards.
02 — How it enforces
Evaluate. Enforce. Escalate.
Three actions, deterministic per policy — configured per agent and per environment.
01
Evaluate in-line
Every candidate response is scored against your policies before delivery, using latency-optimized versions of the same scorers as your offline evals.
02
Enforce deterministically
Pass, block with a fallback, or rewrite-and-retry — explicit actions per policy, versioned like code.
03
Escalate with context
Edge cases route to humans with the full trace attached. Decisions land in the audit log automatically.
NovaGuard is in private beta. Design partners run it on real traffic and shape the policy engine with us — limited seats.
Join the beta03 — Under the hood
What runs in the request path.
A real-time policy engine using latency-optimized scorers at 1:1 parity with offline — deterministic actions, compliance rule packs, and an audit log behind every decision.
Inline policy engine
A real-time engine evaluates each response in-line, before it reaches the user — built to fit conversational latency budgets, not a batch job.
Scorer parity
Latency-optimized variants of the same scorers you run offline — 1:1 parity, so the criteria you evaluate with are exactly the criteria enforced in production.
Deterministic actions
Three explicit outcomes per policy: ALLOW, BLOCK (with a fallback or a rewrite), or ESCALATE to a human queue with the full trace attached — versioned like code.
Compliance rule packs
Prebuilt policy packs for regulated conversations — financial disclosures and healthcare boundaries — drop in rather than author from scratch.
Audit log
Every decision — allow, block, escalate — is audit-logged with its score and reasoning: evidence for auditors, not just a dashboard.
Deploy anywhere
Run it in managed cloud, against your own ClickHouse, or fully on-prem on Kubernetes with Helm — under SOC 2 Type II, HIPAA, GDPR, SSO/SAML, and RBAC.
The loop
Part of one closed loop.
NovaGuard is the runtime arm of the loop: it enforces the same scores NovaEval produces — and what it blocks becomes signal for NovaPilot’s next fix.
Next step
Put your production AI under control.
Start free and ship your first trace in 15 minutes — or book 30 minutes and we’ll integrate live on the call: your stack, your data, your first eval report before it ends.
SOC 2 Type II · HIPAA · GDPR · On-prem & BYO ClickHouse available