# Experiment 001 — Sam Reed interviews Maria Stern **Status:** Not yet run. Awaiting fresh session. --- ## Purpose The first pilot of the method-actor architecture in this project. Tests whether structural isolation between two character contexts — with the coordinator routing rather than generating — produces dialogue that reads more like a real interview transcript than what a single LLM writing both halves would produce. This experiment is also the first run of every supporting piece: the actor system prompt, the orchestration protocol, the event-log schema. Discovering bugs in any of these is part of the experiment. --- ## Characters - **`characters/maria-stern/`** — profile + empty journal. Costa Rican-American marketing director in Mountain View, mid-30s, blended household, early in a long-wanted pregnancy. - **`characters/sam-reed/`** — profile + empty journal. Brooklyn-based freelance journalist working on a long-form piece about late-life parenthood in the Bay Area. --- ## Scene - **`scenes/scene-001-morning-kitchen.md`** — A Wednesday morning, ~9am Pacific. Sam visits Maria's Mountain View kitchen for the interview. Her husband at work, stepsons at their mother's. The scene file contains a per-character perspective section so each actor reads only their character's view of the situation. --- ## Before any actor is spawned — present the two pre-experiment choices to the user Per `CLAUDE.md` and `methodology/how-to-orchestrate-multi-agent-dialogue.md` → *Pre-experiment choices*, the coordinator must present these two questions to the user and capture their answers in `notes.md` *before* spawning anything. ### Choice 1 — How to enforce actor isolation (a) **Sub-agents per actor in the current Claude Code session.** Uses the `Agent` tool to spawn one sub-agent per character. Each spawn is a fresh isolated context. No external API access required. (b) **Python coordinator with separate API conversation IDs.** A `lib/run_dialogue.py` script makes separate Anthropic API calls per actor. Requires `ANTHROPIC_API_KEY` env var. ### Choice 2 — Who plays the coordinator role A. **Parent LLM context** — this session IS the coordinator. Simple, observable, judgement-capable. Carries any context the parent has loaded. Best for pilots. B. **Dedicated coordinator sub-agent.** Isolated, reproducible. Adds an LLM call per turn for a mostly mechanical job. Rarely the right answer. C. **Pure Python code.** Cheapest, most deterministic, zero routing bias. Best for production. ### Recording the user's answers The first two lines of `notes.md` look like: ``` Isolation: (a) — sub-agents in current Claude Code session Coordinator role: A — parent LLM context (this session) ``` Then a free-form section for any additional notes. --- ## How the experiment runs (once the user has picked) 1. **Brief each actor.** Per `methodology/how-to-orchestrate-multi-agent-dialogue.md` → *Per-actor briefing format*, each actor receives, in this order: - The actor system prompt from `actor/system-prompt.md` - Their character's `profile.md` - Their character's `journal.md` - The scene's perspective on the situation for their character (extracted from `scene-001-morning-kitchen.md`) - The instruction: *"The scene begins now. Wait for the other character to speak, or speak first if the scene briefing calls for it."* 2. **Open with the doorbell.** Per the scene file's *Opening* section, Sam arrives at Maria's door at 8:58 and they walk to the kitchen. Coordinator initiates by sending Maria the prompt that the doorbell has rung; she responds; Sam arrives and they begin. 3. **Loop turn by turn** following the protocol in `methodology/how-to-orchestrate-multi-agent-dialogue.md`. Every event (think / speak / route / spawn / correct / journal / close) is recorded in `event-log.json` with timestamps. The clean dialogue is appended to `transcript.md`. 4. **Detect end of scene** per the protocol's end-of-scene rules. Soft cap: 40 turns per character. Coordinator should err toward letting the scene continue, not toward cutting it short. 5. **Journal-write step.** After the scene closes, each actor is asked to write a journal entry in their character's first-person voice. The coordinator appends each to the character's `journal.md`. (Maria's first entry; Sam's first entry — this becomes input to experiment 002.) --- ## Outputs of this experiment ``` experiments/exp-001__maria-and-sam/ ├── README.md ← this file ├── notes.md ← user's pre-experiment choices + post-run observations ├── transcript.md ← the clean dialogue └── event-log.json ← every event timestamped, for chronological replay ``` Plus side-effects: - `characters/maria-stern/journal.md` — Maria's first entry appended - `characters/sam-reed/journal.md` — Sam's first entry appended --- ## How to compare to experiment 002 Experiment 002 (Sam interviews Soo-yeon Park) runs after this one. Same actor system prompt, same coordinator architecture; different character (Soo-yeon vs. Maria); same questioner (Sam, but now with one journal entry of accumulated reflection). The comparison axes: - Does Sam's behaviour visibly evolve between scenes 001 and 002 because of his journal? - Does the dialogue with Soo-yeon look distinctively *Korean older wife* — slow, formal, brief — versus Maria's *Costa Rican-American marketing director* register, even though both actors run on the same model? - Any cross-scene contamination between Maria's content and Soo-yeon's content? (There should be none.)