Data Flow

Employer creates posting
  → Title, JD, competencies (from framework), question spine, depth, mode, style
  → JD embedding generated for talent pool matching

Candidate invited
  → Email sent with single-use, 7-day token
  → Candidate clicks link → privacy → resume (optional) → ready

Interview starts (POST /api/interview/start)
  → Load interviewer rubric from DB (structured or adaptive, based on posting style)
  → Fetch current interviewer version from rubric_versions
  → Build dynamic system prompt:
      Structured: rubric fields + JD + question spine + resume
      Adaptive: rubric fields + JD + competency anchors + weights + resume
  → Create Retell session (voice: web call, text: chat)
  → Pass prompt via retell_llm_dynamic_variables (per-call, no race condition)
  → Stamp interviewer_version on transcript row
  → Mark token as used

Interview runs
  → Retell AI manages conversation using GPT-4.1
  → IntegrityMonitor tracks browser/camera flags in-browser
  → No data stored during interview (transcript is on Retell's side)

Interview ends (POST /api/interview/end)
  → Fetch transcript text from Retell
  → Scrub candidate name → [CANDIDATE]
  → Load scorer rubric from DB
  → Fetch current scorer version from rubric_versions
  → Load competency anchors from competency_framework (by competency_id)
  → Call Claude scorer with: transcript + rubric + anchors + integrity flags
  → Parse JSON scorecard
  → Recalculate weighted_score from competency weights
  → Save scorecard with scorer_version + interviewer_version
  → Update interview_status = 'complete'