Phase 1: POC10 Engine & Feature Parity in React 19
Core EngineClosing the gap from POC8 to POC10 within the existing modular, typed React repository.
Engineering Breakdown Tasks
-
Generalize Targeting Engine: Expand
targeting.tsto support all six capabilities withtaskLibraryFor(capability)and the 3-independent-correct tier progression rule. -
Port 5 Interactive Exercise Views: Implement Cognitive Flexibility (Rule Changed), Creative Thinking, Emotional Intelligence (Stuck on Purpose), Self-Expression, and Systems Thinking using existing
CardandChoiceButtoncomponents. - Port Teacher Portal CRUD: Add child, soft-remove, restore, hard-delete with confirmation, grouping by capability strength/gap, and Sonia demo-only flag.
-
Family & School Context Form: Port 16 parent/teacher questions tagged to capabilities (
cap:logical, etc.). - Automated Test Expansion: Extend the 51 unit tests to cover deterministic scoring and tier transitions across all 6 capabilities.
What the Client Can Test & Verify
Navigate to the curriculum and complete all 6 exercises. Verify that the Cognitive Flexibility sort-rule change triggers at item 2, and Emotional Intelligence retry logs properly.
Add a new student, soft-delete them, restore them, and execute a permanent delete with name confirmation. Verify Sonia remains pinned as demo data.
Submit parent and teacher questionnaires and confirm answers save cleanly per role with correct capability tags.
Phase 2: Production Infrastructure, Database & Edge Gate
Brief Priority 1 & 3Moving from browser localStorage to Supabase persistence, server-side Claude API, and edge password protection.
Engineering Breakdown Tasks
-
Supabase Database Schema: Provision tables for
students,evidence_instances,context_answers, anddelete_audit_logswith foreign keys and Row Level Security (RLS). -
Implement
SupabaseEvidenceStore: Seamless swap from in-browser localStorage to the persistent Supabase backend without altering UI layer code. -
Cloudflare Edge Password Gate (
/api/gate): Single shared password protection with an encrypted HTTP-only cookie. Changeable in the Cloudflare dashboard without redeploying code. -
Secure Claude Proxy Function (
/api/report): Browser never prompts for or touches an API key. Requests proxy through Cloudflare Pages Functions holding the secret key server-side. -
Student Identity Migration: Transition from fragile typed-name matching to persistent
stu_xxxxUUID sessions.
What the Client Can Test & Verify
Open browser DevTools Network tab during report generation. Verify no Claude API key or secret token appears anywhere in request headers or HTML source.
Complete an exercise on Desktop Chrome, then open the student record on Mobile Safari. Real evidence and tier status sync immediately via Supabase.
Open an incognito window. You must enter the shared password before accessing any onboarding or student page. Rotate the password in dashboard to test lock-out.
Phase 3: Milestone & Reporting Engine + Hypothesis Logic
Brief Priority 2 & 4Connecting real exercise evidence to update the student profile report and testing parent/teacher hypotheses.
Engineering Breakdown Tasks
- Milestone Evidence Aggregator: Query service to collect raw evidence instances logged since previous report checkpoint, grouping by capability and filtering flagged attempts.
- Dynamic Milestone Report Generation: Structured prompt service passing empirical tier achievements + observed behaviors to Claude to regenerate an accurate, evidence-backed report.
-
Hypothesis Comparison Engine (Pure TS): Evaluate parent/teacher answers against real exercise observations to calculate
Consistent,Not Yet Testable, orContradicted. - Diplomatic Copy Integration: Inject agreed diplomatic phrasing in Contradicted cases so feedback remains constructive and encouraging.
- Report History & Versioning: Store versioned historical reports in Supabase to track progression over time.
What the Client Can Test & Verify
Complete 3 Cognitive Flexibility exercises to reach Provisional tier. Trigger the Milestone Report and verify that the report updates to quote the real exercise evidence rather than questionnaire guesses.
Fill out the parent form rating Logical Reasoning as high. Intentionally fail or skip the Logical Reasoning exercise, then verify the comparison displays the approved diplomatic Contradicted phrasing.
Phase 4: Child Safety, AI Disclosure & Content Moderation
Brief Priority 5 (Non-Negotiable)Mandatory child safeguards before any real pilot student accesses the application.
Engineering Breakdown Tasks
- Age Verification Gate: Pre-onboarding screen requiring age confirmation before a child enters the questionnaire.
- Transparent AI Disclosure: Unambiguous, child-friendly visual indicator informing the child they are interacting with AI.
- Server-Side Content Moderation: Automated moderation pass on all AI-generated text shown directly to the child (specifically the personal encouragement note).
- Review Flagging System: Auto-flagging for human educator review if attempt duration is under 2s (guesswork) or on 4th+ consecutive attempt.
What the Client Can Test & Verify
Attempt to proceed through onboarding with an unverified or invalid age; verify that the platform cleanly prevents access.
Review onboarding and exercise headers to verify the clear child-friendly AI badge is visible across all interactive views.
Trigger report generation; inspect the server response to confirm all child-facing notes pass automated safety validation before rendering.
Phase 5: Content Scaling & AI Rubric Judgement
Brief Priority 6 (Post-Rubric)Replacing placeholder word counts with Claude LLM-as-a-judge and expanding task variants per capability.
Engineering Breakdown Tasks
-
Server-Side Rubric Judge (
/api/judge): Implement Claude prompt judging Creative Thinking & Self-Expression based on Dr. Chew Lee's formal rubric criteria. - Replace Word-Count Placeholders: Remove the temporary 12-word and 20-word limits with qualitative rubric scoring.
- Task Variant Expansion: Structure task libraries in JSON to support 3+ variants per difficulty tier across all capabilities.
What the Client Can Test & Verify
Submit high-effort creative responses vs generic gibberish; verify the rubric judge evaluates genuine conceptual novelty rather than sentence length.
Run repeated exercise sessions with a student; verify they receive fresh task variants across consecutive attempts.