Dialogue · Experiments

AI generated dialog research

Project documents start here · methodology · system prompts
  • Actor the performer's stable posture; thinking-block rules; closing the scene
  • Coordinator the LLM coordinator's dumb-pipe discipline (scheme 2)
  • Monolithic the single-LLM dialogue generator (scheme 3 — one writer scripting both sides)
  • Auditor the auditor's posture and what to refuse to do
Study 001 — three schemes, same scene

Same characters (Brian Yuan + Shane Lauft), same scene (a Sunday meeting at Red Rock Coffee about a Palo Alto listing and a secondary-share opportunity). Three architectures, each on the model that fits its delivery path: python drives the Anthropic API (sonnet 4.6), subagent and monolithic run inside a Claude Code session (opus 4.7). See the three schemes for the design, the study README for what's being measured, and the cost analysis for what each run actually cost.

Study 002 — Claude vs DeepSeek, same scheme

Same scheme (python), same scene, same characters. Different LLM. Tests whether the framework's discipline transfers across models. See the study README for what's being measured.

After the post-mortem · framework revised
Pilot experiments