Trust & integrity suite — plagiarism/AI detection, proctoring, authorship verification

Why / Context

PRD §7.2C: plagiarism/AI-generation detection and browser-locked assessments protect "the credibility of certificates and mentor time alike" — the certificate's value IS the business as the marketplace scales (PRD §12: AI-shortcut integrity is the core threat to 'verified proof of skill'). Teardown §5.3: project-authorship verification as a headline trust claim is ownable whitespace (AVNI verifies interviews, not authorship; Crio has policy but no buyer-facing claim).

Schema already in place — NO migration required

• IntegrityFlag (subject types incl. ASSIGNMENT_SUBMISSION/MILESTONE_SUBMISSION/QUIZ_ATTEMPT/DEFENSE/CREDENTIAL; kinds PLAGIARISM/AI_GENERATED/SIMILARITY/PROCTOR_VIOLATION/MANUAL_REPORT; status workflow)
• ProctorRecord (quizAttempt 1:1, config/events/verdict)
• Credential.revokedAt/revokeReason + public verify page reflects revocation (MVP M11)
• Quiz.proctored flag; ProjectReview.aiMetadata can carry authorship signals

Implementation plan

1. Similarity pipeline: on milestone/assignment submission, background function computes cross-corpus similarity (MOSS-style token fingerprinting for code via a service like Codequiry/MeasureOfSoftwareSimilarity self-host decision doc) → IntegrityFlag(kind: SIMILARITY, score) above threshold.
2. AI-generation signals: stylometry + commit-history heuristics (single-dump repos, no iteration) + optional detector API; recorded as advisory flags — never auto-fail (documented policy: flags inform mentors/defense, humans decide).
3. Authorship verification claim: combine defense verdict (issue #3) + integrity flags + commit cadence into a per-credential "authorship verified" badge on the verify page + portfolio (the marketable trust claim).
4. Proctored quizzes: browser-lock mode (fullscreen enforcement, tab-blur/copy-paste events, webcam snapshot opt-in per config) writing ProctorRecord.events; verdict review UI for admins; enable per Quiz.proctored for certification exams.
5. Review workflows: admin integrity queue (filter by kind/status/score), resolution actions (dismiss/confirm), confirmed flags → mentor notification, resubmission requirement, or credential revocation (existing revoke flow) per severity matrix.
6. Learner-facing due process: notice of confirmed flags, appeal path (thread in admin queue), policy pages (with issue #24).
7. Metrics: flag rates by course/project, false-positive tracking from dismissals.

Acceptance criteria

• Two near-identical submissions in a project corpus both get SIMILARITY flags with linked evidence
• Proctored quiz records blur/fullscreen-exit events; verdict reviewable; VIOLATION blocks auto-pass pending review
• Confirmed integrity violation can revoke a credential; verify page shows revoked state
• Flags never auto-fail without human confirmation (policy enforced in code paths)
• Credential pages show "authorship verified" only when defense + clean flags criteria met

Dependencies / sequencing

• MVP M9/M10/M11; issue #3 (defense) feeds the authorship claim
• Policy copy coordinates with issue #24

Start by mapping the existing IntegrityFlag, ProctorRecord, Credential revocation, Quiz.proctored, and ProjectReview.aiMetadata flows. Read issue #3 for defense sequencing and issue #24 for policy dependencies. Done means the listed similarity, proctoring, review, authorship, revocation, and human-confirmation acceptance criteria are implemented together.