diff --git a/.planning/PROJECT.md b/.planning/PROJECT.md
new file mode 100644
index 0000000..52abc0b
--- /dev/null
+++ b/.planning/PROJECT.md
@@ -0,0 +1,85 @@
+# MOPC — AI Ranking, Advancement & Mentoring During Rounds
+
+## What This Is
+
+An enhancement to the MOPC jury voting platform that adds AI-powered project ranking after evaluation rounds, an admin dashboard for reviewing/adjusting rankings and advancing projects to the next round, and the ability to assign mentors during non-mentoring rounds (e.g., during document submission or evaluation) with automatic carryover across rounds.
+
+## Core Value
+
+Admins can describe ranking criteria in natural language, the system interprets and ranks projects accordingly, and they can advance the top projects to the next round with one click — all with full override control.
+
+## Requirements
+
+### Validated
+
+<!-- Inferred from existing codebase -->
+
+- ✓ Competition system with ordered rounds (INTAKE → FILTERING → EVALUATION → SUBMISSION → MENTORING → LIVE_FINAL → DELIBERATION) — existing
+- ✓ Jury evaluation with scoring forms and pass/fail criteria — existing
+- ✓ AdvancementRule model with configurable rule types (AUTO_ADVANCE, SCORE_THRESHOLD, TOP_N, ADMIN_SELECTION) — existing
+- ✓ ProjectRoundState tracking per project per round — existing
+- ✓ JuryGroup and JuryAssignment for panel management — existing
+- ✓ CompetitionCategory enum (STARTUP, BUSINESS_CONCEPT) with per-project categorization — existing
+- ✓ Email notification system with Nodemailer/Poste.io — existing
+- ✓ Mentor dashboard route group `(mentor)` — existing
+- ✓ Round engine state machine for round transitions — existing
+- ✓ AI services with anonymization layer — existing
+
+### Active
+
+<!-- Current scope. Building toward these. -->
+
+- [ ] AI ranking engine that interprets natural-language criteria into ranking logic
+- [ ] Admin ranking dashboard with drag-and-drop reordering per competition category
+- [ ] Side panel detail view showing evaluation data for selected project in ranking list
+- [ ] "Advance top X" button to promote selected projects to next round
+- [ ] Admin choice per-batch: send advancement/rejection email OR update status silently
+- [ ] Admin-editable email templates with variable insertion ({{firstName}}, {{teamName}}, etc.)
+- [ ] AI criteria preview mode: admin sees parsed rules before applying
+- [ ] Quick rank mode: AI interprets and ranks directly, admin adjusts after
+- [ ] Mentor assignment during non-MENTORING rounds (evaluation, submission, etc.)
+- [ ] Auto-persist mentor assignments across rounds (unless project eliminated)
+- [ ] Admin override for mentor assignments at any time
+- [ ] AI-suggested mentor-to-project matching with admin confirmation
+- [ ] Notification awareness: warn admin if next round doesn't have auto-emails, so they know to send manually
+
+### Out of Scope
+
+- Award eligibility (Spotlight on Africa, etc.) — separate workflow, later milestone
+- Changes to the juror evaluation interface — already built
+- Real-time collaborative ranking (multi-admin simultaneous drag) — unnecessary complexity
+- Public-facing ranking results — admin-only feature
+
+## Context
+
+The competition is actively running. Evaluations for the first round are complete and the client needs to rank projects and advance semi-finalists urgently (by Monday). The ranking criteria were communicated in a mix of French and English by the organizers:
+
+- 2 yes votes → semi-finalist
+- 2 no votes → not semi-finalist
+- 1 yes + 1 no with ≥6/10 overall → consider as semi-finalist (depending on total count)
+- Special attention to whether evaluations included at least 1 internal + 1 external juror
+
+Categories are STARTUP and BUSINESS_CONCEPT — rankings and advancement happen per-category within a single competition.
+
+The platform already has `AdvancementRule` with rule types but no AI interpretation layer. The `MentorAssignment` concept doesn't yet support cross-round persistence or assignment during non-mentoring rounds.
+
+## Constraints
+
+- **Timeline**: Semi-finalist notifications need to go out by Monday — ranking and advancement are highest priority
+- **Tech stack**: Must use existing stack (Next.js 15, tRPC, Prisma, OpenAI)
+- **Data model**: CompetitionCategory (STARTUP/BUSINESS_CONCEPT) is on the Project model, rankings must respect this split
+- **Security**: AI ranking criteria go through OpenAI — must anonymize project data before sending
+- **Existing patterns**: Follow tRPC router + Zod validation + service layer pattern
+
+## Key Decisions
+
+| Decision | Rationale | Outcome |
+|----------|-----------|---------|
+| AI interprets natural-language criteria rather than hardcoded rules | Client changes criteria between rounds; flexible system avoids code changes | — Pending |
+| Rankings per CompetitionCategory, not per JuryGroup | Categories (Startup vs Business Concept) are the meaningful split for advancement | — Pending |
+| Mentor assignments auto-persist across rounds | Reduces admin work; mentors build relationship with teams over time | — Pending |
+| Admin-editable email templates with variables | Client sends personalized emails in French/English; templates must be customizable | — Pending |
+| Side panel for ranking detail view | Keeps drag-and-drop list compact while providing full evaluation context on demand | — Pending |
+
+---
+*Last updated: 2026-02-26 after initialization*