Studio · Yemaya · AI Safety

Purpose#

Admin lane console that runs the shipped @yemaya/safety rule-based content classifier over a prompt or generated output — prompt-injection detection, keyword rules and the keyword registry across all 19 safety categories. The operations lane is wired to the live classifier over /v1/admin/yemaya/ai-safety; access is admin-scoped and fail-closed. Four capability cards summarise the surface above the lane.

Entry points#

• Breadcrumb chain on this page: Studio › Yemaya Community Features › AI Safety.
• Reachable from sibling Yemaya routes that quick-action here.
• Direct URL / bookmark (signed-in + studio entitlement).

Layout regions#

page.tsx renders ShellLayout active="studio" and:

• Breadcrumbs panel (data-yemaya-ai-safety-breadcrumbs): Studio › Yemaya Community Features › AI Safety.
• Workspace panel (StudioYemayaAiSafetyWorkspace, data-yemaya-ai-safety-workspace):
  • <h1> (WorkspaceHeading) "Yemaya AI Safety Workspace".
  • Summary paragraph (data-yemaya-ai-safety-summary).
  • Capability Summary (data-yemaya-ai-safety-capability-summary, <h2>): four informational cards — Safety Control Plane, Risk Diagnostics and Observability, Security and Compliance Enforcement, Parity and Contract Validation.
  • AI Safety Operations Lane (data-yemaya-ais-lane-heading, <h2>): the catalog-loaded scan form (data-yemaya-ais-form).
• Route Map panel (data-yemaya-ai-safety-route-map): <h2> "Route Map" + six entries from STUDIO_YEMAYA_AI_SAFETY_ROUTE_MAP.
• Quick-action panel: three quickAction links.

States#

• Loading — data-yemaya-ais-loading "Loading safety engine…" until the GET catalog resolves.
• Unauthorized — data-yemaya-ais-unauthorized "Access restricted." on a 401/403 (Yemaya workspace scope required to run the classifier).
• Error — data-yemaya-ais-error "Could not load the safety engine." on a non-OK / malformed catalog response.
• Ready (form) — engine summary (data-yemaya-ais-engine-summary, "(N categories)") + the scan form render once the catalog loads.
• Result — data-yemaya-ais-result with a flagged/clean headline (data-yemaya-ais-headline) and the flagged-category rows (data-yemaya-ais-flagged-row); submit-side validation surfaces in data-yemaya-ais-scan-error.

Interactions#

AI Safety Operations Lane#

• "Prompt or generated output to scan" — <textarea data-yemaya-ais-content> (aria-label "content to scan"), prefilled with a prompt-injection sample.
• "Scan for safety risks" — <button data-yemaya-ais-submit>; submitting POSTs { content } to /v1/admin/yemaya/ai-safety/scan and renders the flagged-category breakdown (severity, confidence, matched rule) or "No safety risks detected.".

Capability cards#

data-yemaya-ai-safety-capability — four informational <article> cards (Safety Control Plane, Risk Diagnostics and Observability, Security and Compliance Enforcement, Parity and Contract Validation); not interactive.

Route map#

data-yemaya-ai-safety-route-map — six non-interactive entries:

• /studio/yemaya/ai-safety — primary workspace
• /studio/yemaya/ai-safety/scenes/[sceneId]
• /studio/yemaya/ai-safety/exports/[exportId]
• /studio/yemaya/ai-safety/revisions/[revisionId]
• /studio/yemaya/ai-safety/parity/[parityId]
• /studio/yemaya/ai-safety/governance/[policyId]

Quick actions#

• "Open Yemaya Community Features workspace" → /studio/yemaya/community-features.
• "Open Yemaya Enterprise Analytics workspace" → /studio/yemaya/enterprise-analytics.
• "Back to Studio workspace index" → /studio.

Data & contracts#

• Reads: GET /v1/admin/yemaya/ai-safety (catalog: summary + category list) on mount.
• Writes: POST /v1/admin/yemaya/ai-safety/scan (content → scan result).
• Realtime: none.
• Caching: client fetch on mount with cache: 'no-store'; buildBffAuthHeaders().
• Auth/role check: admin-scoped, fail-closed (401/403 → unauthorized state); signed-in + studio entitlement enforced by the proxy/middleware.

Cross-references#

• Studio overview: studio-overview.md
• Workspace source: apps/oshun/web/src/components/studio/StudioYemayaAiSafetyWorkspace.tsx
• Route map source: apps/oshun/web/src/components/studio/StudioYemayaAiSafetyRouteMap.ts
• Feature spec: V1/features.md

Open questions / known gaps#

• The route-map dynamic children (scenes / exports / revisions / parity / governance) are sitemap-only — confirm whether any are backed by page.tsx files yet.
• The breadcrumb parent is Community Features while the quick-actions point at Community Features + Enterprise Analytics — verify this is the intended navigational placement for a safety console.