Skip to main content
This page captures the major Eval Labs hardening steps that made the system ready for real Lucia development.

​
April 2026 milestone

Eval Labs became usable for active Lucia development after the custom prompt, review, tester identity, environment, CORS, and persistence work landed.

​
Key product changes

​
Custom prompt launcher

Commit: 
884f194 evals: add custom prompt launcher and saved suites
Added:
  • fork landing page
  • custom 1–10 prompt launcher
  • saved custom suites
  • custom run history
  • runSource: custom
  • shared Review Queue reuse

​
Review completion and navigation

Commit: 
3ef6fb5 evals: improve review completion and navigation
Added:
  • final prompt button changes from Save & Next to Save
  • completion action area
  • top-left brand home navigation
  • clickable breadcrumbs
  • dist/ removed from Git tracking

​
Tester identity exports

Commit: 
16f53cd evals: attach tester identity to exports
Added:
  • TesterIdentity
  • prompt-level savedBy
  • top-level exportedBy
  • reviewer identity in CSV/Markdown exports
  • Clerk identity normalization

​
Supabase run item conflict target

Commit: 
d51f9b1 evals: fix run item upsert conflict target
Changed initial eval_run_items upsert target from id to run_id,item_index.

​
Supabase row identity reconciliation

Commit: 
fd8a366 evals: reconcile run item ids before upsert
Fixed primary-key collisions by reusing existing run item row IDs for the same logical run_id + item_index slot.

​
Environment hardening

​
Netlify

Changed Eval Labs endpoint to: 
https://api-dev.hellolucia.ai/admin/operator-focus

​
Render dev Engine

Updated ADMIN_ALLOWED_ORIGINS to include: 
https://evaluationlabs.ai
https://www.evaluationlabs.ai

​
Validation outcome

Validated:
  • Eval Labs deployed site calls api-dev
  • Engine returns 200
  • custom prompt run succeeds
  • Supabase persistence succeeds
  • no CORS error
  • no eval_run_items 409 after latest bundle
  • exported identity metadata works

​
Current status

Eval Labs is ready for active Lucia dev refinement and has passed the AI-reviewed platform readiness gate.
Use custom prompt suites as the primary tool for behavior-family refinement. Do not overclaim this as human Lucia-quality approval.

​
May 2026 review-layer milestone

Eval Labs gained a full layered review architecture:
  • adjudication-ready review schema
  • guided Employee Review fields
  • suggested review layer
  • Human Guidance Evaluation
  • Quick Review UX for non-expert reviewers
  • review state controls and routing
  • adjudication queue filters
  • canon-candidate workflow
  • JSON, CSV, and Markdown export parity for structured review evidence
  • lifecycle finalization
  • Supabase promptRecord payload persistence
  • dirty/completion state preservation
  • semantic stepped rating sliders
  • native-feeling confidence bar visual design
Current doctrine: 
Employee reviewers capture reaction.
Senior adjudication assigns canonical meaning.
This should be treated as a major product and Canon milestone, not a cosmetic UI change.

​
May 2026 product-surface and access milestone

Eval Labs was refined into a more complete internal product surface:
  • top app shell owns page identity
  • in-page blog-style mastheads were removed from the app
  • Custom, Auto-generated, and Controlled Batch Runner surfaces were split
  • /lucia/auto-generated became the canonical normal generated tester route
  • /lucia/automated remains a legacy alias
  • /analysis became the canonical Global Analysis route
  • /experiments remains a legacy alias
  • Single Run Analysis was added at /analysis/runs/:sessionId
  • Run rows were standardized with two-zone layout and Copy dropdown patterns
  • Copy Session ID / Copy Deep Link controls were added across key surfaces
  • Global Analysis loading was fixed to show immediately
  • role-gated owner/admin/evaluator behavior was added as the initial product gate
Historical access limitation at this milestone: 
Backend/RLS enforcement still required before external evaluator rollout.
Current role and RLS posture is documented in Current System State and Eval Labs Roles and Access Matrix.

​
May 2026 AI-reviewed platform readiness gate

Final gate result: 
60 / 60 completed runs
3,000 expected prompts
3,000 eval_run_items
3,000 Lucia responses
3,000 reviews
Supabase verification result: 
ready | 60 | 3000 | 3000 | 3000 | 3000
localStorage verification: 
sessionCount = 60
persistedLocalFullPayloadSessionCount = 0
persistedLocalHasItemLevelData = false
persistedLocalItemLevelDataSessionCount = 0
ownedSessionCount = 60
otherOwnerSessionCount = 0
ownerlessSessionCount = 0
rawByteSize around 68,815
This proves:
  • run creation
  • Lucia response capture
  • review generation
  • review persistence
  • run finalization
  • Run History truth
  • Global Analysis truth
  • Supabase/UI count agreement
  • localStorage compactness
  • controlled batch lifecycle
  • no visible cross-owner local leak in the tested owner context
This does not prove:
  • Lucia is human-approved
  • Lucia is ready for real operator use
  • employee rollout is complete
  • human evaluators agree with AI scoring
  • backend/RLS permissions are complete by themselves
Read next: AI-Reviewed Platform Readiness Gate.