FEATURE SPEC¶
Guardrail Policy & Configuration Console¶
Novaworks AI | x_novaw_guardrails | Stage: Refine | DRAFT
Parent Feature Set: ../concept.md.
History¶
| Version | Date | Author / Notes |
|---|---|---|
| 1 | July 29, 2026 | Eswar Vandanapu — Initial draft, derived from PRD Feature 3. |
Problem Statement¶
Customer SN Administrators have no way to tune sensitivity, thresholds, or which risk categories route to a human, per customer or per pilot group. Guardrail policy today would otherwise be a single hardcoded, global setting.
Scope¶
Let an administrator tune: - Which checks run - At what threshold - Which risk categories route to a human
Customer-wide by default, with group or rule-based overrides for piloting a stricter or looser policy. This builds on the same targeting pattern as the platform feature-flag service rather than introducing a second mechanism.
Out of Scope¶
- The underlying input/output checks themselves (see ../input-guardrail-engine/spec.md, ../output-guardrail-engine/spec.md)
- The sensitive-conversation classifier itself (see ../sensitive-conversation-detection-routing/spec.md) — this console configures its thresholds/routing, not its model
Value¶
A quick win per the PRD's grouping — extends the existing feature-flag targeting layer rather than building a second configuration mechanism, giving administrators a familiar console to pilot stricter or looser guardrail policy per customer or cohort.
Delivery¶
Extend the feature-flag targeting layer to carry guardrail thresholds and routing rules, and give administrators a console to set and pilot them.
Acceptance Criteria¶
- [ ] Feature-flag targeting layer is extended to carry guardrail thresholds and routing rules
- [ ] Administrator console lets an admin configure which checks run, at what threshold, per customer
- [ ] Administrator console lets an admin configure which risk categories route to a human
- [ ] Group/rule-based overrides are supported for piloting a stricter or looser policy than the customer-wide default
- [ ] Console changes take effect without requiring a deploy (same operational model as feature flags)
Dependencies / Related Features¶
- Nova Platform Enhancements PRD v0.4: this feature extends that PRD's feature-flag/targeting service rather than building a new mechanism.
- Feature 1 — Input Guardrail Engine and Feature 2 — Output Guardrail Engine: consume the thresholds/checks this console configures.
- Feature 5 — Sensitive Conversation Detection & Routing: consumes the routing-category configuration this console sets.
Risks¶
| Risk | Severity | Mitigation |
|---|---|---|
| Misconfiguration silently weakens guardrails for a customer/cohort | Medium | Console changes should be captured in the Feature 4 audit log so a policy change is itself traceable. |
| Divergence between feature-flag targeting semantics and guardrail-specific needs (e.g. threshold ranges, category enums) | Medium | Confirm the targeting layer's data model can express guardrail-specific fields before extending it. |
Open Questions¶
- Figma link / UI design for the console is not yet available.
- Does extending the feature-flag targeting layer require a schema change to that service, and if so, is that owned by this Feature Set or the platform enhancements track?
Status: DRAFT — derived from PRD Feature 3. For internal engineering review only.
Novaworks AI | Confidential