Trivia & Tunes · CaveauAI
Editable Mermaid sources for the four flows behind a live game: the per-question AI pipeline, the muse recall/remember loop, the graded-dataset accumulation, and the flywheel's next turn.
From a player's spoken answer to the host's reply — every model named.
flowchart LR
A[Player answers aloud] --> B["Whisper STT (medium)"]
B --> C[Grading: Azure GPT-4o-mini]
C -->|low confidence| C2[Fallback: Nova-lite]
C --> D[Verdict + score 0-10]
D --> E[Enrich: trivia-facts RAG]
E --> F["Host line: Claude Haiku (Bedrock)"]
F --> G[TTS: ElevenLabs or Piper - Finn speaks]
C --> T[(ai_usage_events: cost telemetry)]
F --> T
host-muse corpus. The host remembers the room across questions and across nights.
flowchart TB
M[Moment: round_intro / reveal / standings] --> R{tt_muse_recall}
R -->|build query from context + host id| S[search: host-muse corpus]
S --> F[Filter to this host tag, max 320 chars]
F --> P[Inject 1-2 memories into host prompt]
P --> L[Host speaks a callback]
L --> W{tt_muse_remember}
W -->|reveal / standings only| D[ingest: save payoff moment]
D --> M
Every finalized game ships to the accumulating qa-history corpus.
flowchart LR
G[Game session ends] --> Q[ai_grade_log: finalized rows]
Q --> X[exportSessionToCorpus]
X --> Y[One qa-history document per game]
Y --> Z[(CaveauAI corpus · chunked + embedded)]
Z --> H[Future grading recalls precedent]
Z --> K[Question-bank enrichment]
H --> G
Honest roadmap: today's signal becomes tomorrow's sovereign model.
flowchart LR
A[Accumulated qa-history + routing telemetry] --> B[Predictive difficulty model]
A --> C[Sovereign fine-tune dataset]
B -.coming next.-> D[Auto-calibrated question sets]
C -.coming next.-> E[Private fine-tuned host model]
D --> F[Better games]
E --> F
F --> A
Mermaid now; the best of these are redrawn as bespoke SVG for publication.