Every figure Viva publishes — a session verdict, a credential, a skill tag — is produced by the closed loop described on this page. This document is the citable reference for schools, admissions offices, and employers.

The closed loop

• Submission intake. The student's own work (essay, report, code, poster) is ingested and content-extracted with quality gates.
• Forensic pre-read. An AI examiner maps the submission: claims made, evidence used, places where a non-author would fail.
• Personalized question bank. Questions are generated from the submission itself — no generic quizzes. Instructor custom questions are asked verbatim.
• Live adaptive oral defense. A real-time voice interview. Follow-ups adapt to each answer; reading aloud and coached answers alter the questioning.
• Identity & integrity checks. Voiceprint match against the student's enrolled baseline, visual probes, focus monitoring, and coaching detection run during the session.
• Multi-dimensional grading. The transcript is scored for Understanding and Authenticity, with fairness guardrails. Scores are suggestions an instructor can override.
• Sealed record. The session becomes a tamper-evident record; aggregates can be issued as signed, verifiable credentials.

What is scored

• Understanding. Can the student explain the why behind their own work — reasoning, implications, trade-offs — under live follow-up questioning?
• Authenticity. Is the oral account consistent with the written submission? Contradictions, unfamiliarity with one's own terms, and fabricated process memories lower it.

Results are expressed in four tiers — Excellent, Strong, Proficient, Developing. Numeric scores exist for instructors; every AI-suggested grade is instructor-overridable, and schools may hide numbers entirely.

The integrity stack

• Voiceprint verification. Speaker embedding compared against the student's enrolled voice baseline, per language, across the whole session.
• Visual probes. Session-specific images the student must discuss — unanswerable by an off-screen helper who cannot see them.
• Focus & camera checks. Optional tab-switch monitoring and on-device camera analysis (no biometric images leave the browser).
• Coaching detection. Speech-pattern divergence from the student's own baseline flags real-time AI or human coaching.
• Fairness guardrails. Accent and ASR variants of a student's own terms are never treated as unfamiliarity; approved accommodations suppress the related signals.

Reliability: the grading council

Selected interviews are independently re-scored by two additional AI models from different providers, blind to the original grade. An interview is selected when it is flagged for review, sits near a grade boundary, belongs to an instructor-designated high-stakes assessment, or when an instructor requests a check on demand.

• Bracketing consensus. The independent models sit on either side of the primary grader by design. When the primary grade falls between them, the grade is corroborated.
• Directional signal. When both independent models land on the same side, the report says so — "second reads lean higher / lower" — an actionable review signal, never an automatic change.
• Advisory only. The council never alters a stored grade. Instructors keep final judgment.

Verifiable credentials

Students may issue two kinds of credentials from their verified record — a Subject Credential (one class over a term) and a Communication & Reasoning Profile (cross-subject). Both are aggregates only: no transcripts, no per-assignment scores.

• Earned, not claimed. A credential can only be issued from completed, integrity-checked oral defenses. No record, no credential.
• Signed. Each credential's payload is signed (HMAC-SHA256). Any alteration fails verification.
• Student-owned. The student issues it, shares it by unguessable link, and can revoke it at any time. Reissuing keeps the same link.
• Skill tags with evidence. Skills shown on a profile carry an evidence count — how many separate live defenses demonstrated that skill.

Verification API for institutions

Any institution can verify a credential live — no account, no PDF trust:

GET /api/credentials/verify?id=<credential-id>

The response reports validity, revocation, the signature, issue dates, and the attested aggregates. A tampered or revoked credential fails immediately.

Data commitments

• Public surfaces show aggregates the student chose to publish. Raw transcripts and recordings are never public.
• Credentials and profiles are shared only by their owner and are revocable at any time.
• Viva does not sell student data. Verification is free for the institutions that check it.

Questions about this standard, institutional verification, or accreditation alignment: hello@vivaproof.com
Also see: Viva Card & verification · Trust Center.