Skip to main content
Eval Labs now has separate surfaces for owner/admin oversight, evaluator workbench testing, tester prompt-testing, guest-facing verification, controlled platform checks, derived Registry Diagnostics, persisted Behavioral Observatory labels, Run History, Team Review, Global Analysis, Single Run Analysis, and Review Queue work.

​
Route map

​
Surface definitions

​
Role-aware home

The home route is role-aware. Owner/admin see the privileged platform overview. Evaluator and tester users should see onboarding/workspace entry points appropriate to their role.

​
Launcher

The Launcher is the workspace chooser. It separates:
  • Custom Prompt Test
  • Auto-generated Prompt Test
  • Guest Facing Agent Verification Check
  • Controlled Batch Runner
The top app shell owns page identity. The older in-page blog-style masthead pattern has been removed from the product surface.

​
Custom prompt tester

The Custom Prompt Test lets a user enter 1-10 exact prompts. Use it for targeted testing, evaluator work, and repeatable behavior-family review. Owner/admin, evaluator, and tester roles can use this surface.

​
Auto-generated prompt tester

The Auto-generated Prompt Test runs the normal generated 50-prompt test. Owner/admin, evaluator, and tester roles can use this surface. It is separate from controlled batch infrastructure.

​
Guest Facing Agent Verification Check

Guest Facing Agent Verification Check runs the booked-guest verification scenario pack through the app surface. Owner/admin and evaluator roles can use this surface. Tester cannot use this surface.

​
Verification Results

Verification Results shows the saved/current Guest Facing Agent verification output, scenario failures, exports, and copied summaries. Owner/admin and evaluator roles can use this surface. Tester cannot use this surface.

​
Controlled batch runner

The Controlled Batch Runner is controlled platform-readiness tooling. It supports:
  • 1-run smoke
  • 3-run checkpoint
  • 10-run checkpoint
It was used for the 60-run readiness gate. Owner/admin and evaluator roles can use it in the current role model. Tester cannot use it.

​
Run History

Run History is the scoped run ledger. It includes completed/finalized run truth and may show scoped operational state. Owner/admin can inspect shared/global persisted evidence. Evaluator and tester access is scoped to their own allowed work.

​
Team Review

Team Review is the owner/admin oversight surface. It groups evaluator activity, flags review gaps, and helps owner/admin decide where human evaluation signal needs inspection. It is not available to evaluator, tester, or unassigned users.

​
Registry Diagnostics

Registry Diagnostics is the read-only diagnostic surface for the Dataset Registry and Human Review Queue 2.0 classification model. It shows derived suggestions from existing Eval Labs run/review data:
  • dataset membership suggestions
  • confidence
  • source fields
  • queue lane suggestions
  • Noise / Watch classification patterns
It does not create labels, save human behavioral decisions, or prove final dataset membership. /registry-diagnostics is canonical. /dataset-diagnostics remains a legacy inbound alias.

​
Behavioral Observatory

Behavioral Observatory is the first-class behavioral labeling surface. It lets a reviewer inspect conversations and save structured labels:
  • Intent
  • Guest Affect
  • Response Strategy
  • Humanness
  • Notes
Saved labels are persisted Behavioral Observatory evidence only after Supabase confirms the save. Derived suggestions on this page are starting context, not final judgment.

​
Global Analysis

Global Analysis is the read-only owner/admin behavioral and analytics surface. It is AI-analyzed platform evidence, not human Lucia-quality approval. /analysis is canonical. /experiments remains a legacy alias.

​
Single Run Analysis

Single Run Analysis is read-only analysis of one completed run/session. It includes run metadata, behavioral summaries, item rows, and copy/deep-link controls when hydrated data is available.

​
Review Queue

The Review Queue is the scoring and review workflow for prompts/items. In the current role model, evaluator and tester review access is scoped to their own allowed work. Owner/admin can inspect shared persisted evidence where oversight applies.

​
Copy controls

Copy Session ID and Copy Deep Link controls exist across key surfaces, including run rows, controlled batch summaries, Single Run Analysis, and review/item contexts. These controls are addressability infrastructure. They make future debugging, review handoff, and Canon recovery easier.

​
Surface distinction rule

Use this map:
Do not claim a diagnostic suggestion is a saved label.