Human + AI characters — interactive panel

Design decisions locked in this session

engine

megaprompt only, for now — one model voices the whole cast. Most efficient, and the most flexible for changing format, cast size, or deliberation depth. (Python / isolation stays the batch thesis, deferred.)

the cast

starts as three — Li Ka-shing (investor), Vladislav Doronin (developer), Shane Lauft (realtor) — plus the human.

live cast

add or retire characters at any turn boundary. Add appends materials (a one-time cache write); retire is a parting line + a "stop voicing X" instruction (materials stay cached, dormant). Joiners are omniscient. Keep ≥1.

the human

a real user, not an actor — here a Shanghai buyer choosing among three properties (the exp-009A case). Strict in-world; identity undefined / adaptive.

opening

each character self-introduces briefly (own voice); the last invites the guest to say who they are and what they're deciding. No MC.

deliberation

before responding, the characters confer backstage — a flowing 3–4 exchange debate (no numbered rounds), shown in a slide-out side panel. No separate private <think>; the one model stages it all.

responses

one per character, relevance-led (the most-implicated leads); each follower answers the line actually delivered. Distinct, still disagreeing. @name → direct answer, confer skipped.

cost

each turn records its tokens + API cost — shown left of the floor marker, with a running session total in the sub-strip. Turn 1 is the dear cache-write; later turns are cheap cached reads.

controls

the Send button doubles as ■ Stop — while the panel is talking, pressing it interrupts and takes the floor back (replaces the separate "jump in").

artefacts

standard event-log.json + transcript.md on close (now with deliberation + per-turn cost events) → replays in the existing viewer, feeds the audit.

backend

local Python SSE server (not CF Pages); only the recorded session is published.