Changelog

Knowledge, Trust and Evidence layers publishedTransparency

• Knowledge Center: About, How It Works, Why Not Just ChatGPT?, a canonical reference for AI systems, the Founder Badge, and a 139-question FAQ.
• Educational Knowledge Hub: 12 free guides teaching exam-mathematics method, with no sign-up required.
• Our Educational Principles: the six positions the platform is built on, and how each changed the software.
• Trust Center: an explicit limitations list, guidance on when a human teacher is the better choice, and a deliberately empty student-stories section explaining why nothing was invented to fill it.
• Evidence Center, architecture transparency, this changelog and a public roadmap.
• A CI gate (scripts/validate-knowledge-layer.mjs) that fails the build on knowledge contradictions, fabricated social proof, or a claim drifting from its source of truth.

AI Economics — cost and coverage analyticsPlatform

• A service catalogue and model-call telemetry layer, so every AI operation is attributable to a cost.
• A cost engine and economics layer covering per-lesson economics, student consumption, credit summaries and service quality.
• Owner analytics delivered across milestones M1–M4.3, each with its own engineering review, release gate and closeout document.

Not a student-facing feature. It is here because sustainable unit economics is what allows the free tier to stay genuinely useful, and because the platform documents its internal work to the same standard as its public work.

AI failure observabilityReliability

• Structured recording of AI tutor failures, so a student-visible error becomes a diagnosable event rather than a support ticket.
• Telemetry access RPCs for the AI Monitor.
• Grant hardening applied to the new surfaces at the same time they shipped.

Security hardeningSecurity

• A full production security audit across the frontend, database, Edge Function, authentication and deployment configuration.
• Privilege-escalation and billing-bypass paths closed; the AI knowledge base locked against cross-user writes.
• Security headers and a Content Security Policy added; CDN dependencies pinned with subresource integrity.
• RPC grant hygiene tightened.

The audit report is maintained in the repository, including the findings that were rated low severity and the recommendations still outstanding.

Operation-based credits and the Study PlannerPlatform

• Credits moved from a flat per-message charge to per-operation pricing, so a student using cheap operations no longer subsidises one using expensive operations.
• A race condition in AI credit refunds fixed.
• Study Planner persistence.

Taxonomy v1 — the diagnostic vocabularyDiagnosis

• A fixed taxonomy of 5 topic domains and 33 skills with permanent identifiers, applied across question records, weakness signals, mastery records, weakness reports and focus tasks.
• Unmapped AI detections logged rather than silently accepted, so gaps in the taxonomy surface instead of corrupting the diagnosis.
• Every stored record now carries the taxonomy version that produced it.

Arguably the most consequential change on this list. Before it, the same weakness could be recorded under three different names and no pattern would ever be visible.

Adaptive Verification — Phase 0Diagnosis

• Verification columns added to question records, the groundwork for distinguishing a confidently correct diagnosis from an uncertain one.

Weakness Analyzer and Focus Practice matureDiagnosis

• Severity bands, recency weighting and trend added to weakness reports; weakness signals linked to the specific question that produced them.
• Signal-aware focus plans with a proper lifecycle and single-active-plan enforcement.
• Idempotency for exam sessions and question records, so a retried write cannot duplicate or corrupt a student's history.
• Daily streak tracking.