AI first-pass project review + mentor rubric drafting

Why / Context

PRD §7.2A calls AI first-pass project review "the single highest-leverage move for scaling the mentor-guided project engine" — it multiplies effective mentor throughput without hiring more mentors, directly attacking the mentor-supply bottleneck listed in PRD §12 Key Risks. Paired with AI-assisted rubric drafting (mentors edit an AI draft instead of writing from a blank page), it raises how many learners one mentor can serve profitably.

Schema already in place — NO migration required

• ProjectReview(kind: AI_FIRST_PASS, reviewerId: null, aiMetadata Json, summary Json)
• RubricScore rows attach to any ProjectReview — AI proposes, mentor confirms
• MilestoneSubmission.status=IN_REVIEW gating
• AiUsageEvent(feature: FIRST_PASS_REVIEW | RUBRIC_DRAFT)
• IntegrityFlag(kind: AI_GENERATED|SIMILARITY) for suspicion signals surfaced during review

Implementation plan

1. Trigger: on MilestoneSubmission creation for review-checkpoint milestones, enqueue AI first-pass (Netlify background function) before the mentor queue entry becomes actionable.
2. Review harness (src/lib/ai/first-pass.ts): fetch submission artifacts (repo URL → clone summary/tree + key files within limits, notes, prior feedback), evaluate against Rubric.criteria + Project.heldOutEvalConfig descriptors; output per-criterion provisional scores + evidence quotes + revision suggestions. Claude API — check model ids/pricing via the claude-api skill at build time.
3. Persist as ProjectReview(kind: AI_FIRST_PASS, status: COMPLETED) with aiMetadata (model, tokens, confidence) and provisional RubricScore rows.
4. Mentor queue UX: submission card shows AI verdict chip; mentor review form pre-fills AI scores/comments as an editable draft (RUBRIC_DRAFT usage event on regeneration).
5. Fast-fail path: if AI confidence is high that the submission is incomplete (missing deliverables), auto-return CHANGES_REQUESTED with checklist — configurable per project, default off (mentor approves the auto-return for the first N).
6. Sprint tier integration: AI first-pass is the ONLY review — auto-approve above threshold (issue #2).
7. Quality loop: mentor agreement tracking (AI score vs final mentor score delta) on the admin AI dashboard; tune prompts per category.
8. Meter everything via the shared wallet/cost pipeline from issue #4 (platform-paid, not learner wallet).

Acceptance criteria

• Every checkpoint submission gets an AI_FIRST_PASS review before mentor action; mentor sees it as a pre-filled draft
• Mentor can accept, edit, or discard the draft; final MENTOR_* review is always the decision of record
• Sprint submissions auto-complete on AI pass without mentor involvement
• AI-vs-mentor score deltas visible in admin; usage metered per review
• Repo fetch respects size limits and never executes untrusted code

Dependencies / sequencing

• MVP M10 (review loop) required; issue #4's cost/metering plumbing shared
• Blocks the Sprint auto-review path in issue #2

Start by reading the MilestoneSubmission review-checkpoint flow and src/lib/ai/first-pass.ts, then trace the shared cost and metering pipeline from issue #4. Confirm the sequencing with MVP M10 and issue #2 before implementing the review, mentor draft, sprint, and admin agreement paths. Done means checkpoint submissions receive safe, metered AI reviews while mentor decisions remain authoritative and repo fetching never executes untrusted code.