Each experiment is a single dialogue scene. The 4-window viewer shows what each actor was thinking (private) and saying (public) in real time, alongside the coordinator's routing log and the clean transcript. See README.md for the project shape, or methodology/ for the design rationale.
A long-form journalist interviews a Costa Rican-American marketing director, eight weeks pregnant after two prior losses, in her Mountain View kitchen. Tests whether structural isolation produces dialogue more faithful than single-LLM multi-character generation. The room changes once around I don't usually sit and the half-dead willow she'd walked past two hundred times.
The same journalist, by Zoom, interviews a Korean upper-middle wife in Seoul about transnational families. Same actor system prompt; new character; Sam now has a journal entry from exp-001 carrying forward two technique self-corrections. Tests whether structural isolation holds across distinctly different character registers — and whether Sam's journal-mediated learning shows. The closing sentence, Nobody asks me that, is the kind her profile predicted.
The pilots above surfaced four classes of failure: cross-perspective word leakage, named-technique over-deployment in thinking, third-person POV slips, and locked-fact violations in thinking. The profile, scene, and orchestration methodologies, the actor system prompt, and the audit procedure were rewritten. The experiments below run under the revised framework.