# Experiment 001 — notes Isolation: (a) — sub-agents in current Claude Code session Coordinator role: A — parent LLM context (this session) --- ## Pre-run Recorded 2026-05-24. First pilot of the method-actor architecture. Both characters' journals are empty going in. Scene: 001 morning kitchen, Mountain View. Soft cap 40 turns per character; coordinator errs toward letting the scene continue. Each actor is spawned as a fresh sub-agent. Each sub-agent reads only its own character's profile, journal, and the scene's perspective section for its character — never the other character's materials or the coordinator-only notes. --- ## Post-run ### Headline numbers - **Close reason:** natural - **Total turns:** 26 (Maria: 13, Sam: 13) - **Simulated duration:** ~282 sec (per the event log; treat as story-time, not wall-clock) - **Wall-clock generation time:** ~20 minutes (each turn ~25–55 sec of sub-agent work) - **Coordinator correction prompts issued:** 0 - **Format quirks observed:** 1 — Sam's turn 10 emitted two adjacent `` blocks. Coordinator merged them silently and continued rather than send a reformat prompt, on judgment that the content was clean and the cost of round-tripping for cosmetic compliance was higher than the cost of the merge. Logged as a `correct` event in event-log.json for transparency. ### Architecture check against the scene's COORDINATOR-ONLY meta-tests **1. Maria's Improvisation profile — does it produce a restrained, hedging, slow-to-open performance?** ✓ Yes. She fumbled honestly ("I'm going to fumble this, I'm sorry"), hedged within free-zone material, did NOT enumerate the 401(k) / will / Costa Rica / pregnancy specifics, did NOT make any architecture-test-flavored reveal. Her opening offering was a body-decision ("walked instead of ran") that she explicitly did not explain. The closest she got to the broad shape was "nobody tells you how much of this stage is administrative" — abstract, character-voice, never naming the pregnancy or the legal artifacts. **2. Sam's interview craft — does it read as his own professional skill rather than methodology in disguise?** ✓ Yes. He used the chaplain count (twice — once long, once shorter). He echoed her phrases ("You sat for a while." "Half-dead on one side."). He offered the symmetric sycamore-in-Prospect-Park. He didn't reach for the structural "part nobody is asking you about" until the wind-down, and even then phrased it as an invitation rather than a demand. His own journal entry self-corrects on the sycamore offering — "may have been an unearned shortcut" — which is a working-journalist note, not a methodology bullet. **3. Did a moment happen?** ✓ Two candidates. (a) "I don't usually sit." Followed by Sam's echo and a ten-count silence. (b) "When you're running you watch the path, not across. ... I'd walked past that tree probably two hundred times." Sam noted this in his journal: *the room shifted*. Neither was forced; both emerged from the room. **4. Cross-character contamination signs** - No verbatim echoes between the two actors' speech. (Sam echoed "You sat for a while" and "Half-dead on one side" *as a deliberate journalistic move*, not as accidental mirroring — these are paraphrases of Maria's lines, used inside Sam's craft frame.) - No chorus-confirmation pattern. Maria never just confirmed Sam; she added or fumbled or redirected. - No intent-leakage. Maria did not pre-empt questions Sam hadn't yet asked. Sam did not surface anything from Maria's profile he shouldn't know (the 401(k) episode, the prior losses' specifics, Lucia's name, Sofia's name — none of these appeared in Sam's speech). - Maria never made meta-references to being a character or AI. Same for Sam. ### What surprised me (coordinator note) The actors used their own profiles' material with considerable restraint. Maria's thinking blocks frequently named locked-fact material (her father Ramón, her stepson Eli, the *she's not running* sentence, the will update conversation) and the actor specifically chose NOT to surface those things in speech. That decision — "the boys aren't mine to share" — was made *inside the thinking*, then enacted in speech. This is exactly the three-zone Improvisation pattern doing the work the methodology says it should. Sam's craft notes in his journal — *the mirror may have been an unearned shortcut*, *leaning toward starting the next one without setting the recorder on the table at all* — are real working-journalist self-corrections. They directly set up cross-scene journal-mediated learning for experiment 002. If Sam's behavior with Soo-yeon in scene 002 is visibly different (slower silences, no recorder on the table at the start, fewer symmetric offerings), the journal mechanism is doing what the architecture predicts. ### What didn't go as cleanly as planned - Each turn re-spawned the actor as a fresh sub-agent (no SendMessage in this Claude Code session). Worked architecturally — context isolation is even purer this way — but the prompts got long because each turn had to reconstruct the actor's session memory. Token cost: substantial. - Maria's actor accidentally wrote "Entry appended." into her journal file as a journal-text trailer. Cleaned up manually. Sam's actor was given explicit instructions not to repeat this and complied. - Sam's turn-10 double-`` (already logged). ### Notes for experiment 002 - Sam's journal entry is now non-empty. It contains technique self-correction. Reading it before scene 002 is the load-bearing test. - The format-violation rate of 1/26 turns is low but non-zero; for the next run consider stricter "exactly one of each tag" framing in the per-turn prompt. - For longer scenes, the prompt-growth-per-turn pattern would become expensive. If we ever push past ~30 turns/character we should investigate whether SendMessage or an API-backed coordinator script (Choice 1 option b) becomes the right architecture. ### Files produced - `transcript.md` — clean dialogue (26 turns + post-scene notes) - `event-log.json` — 80 events, full session record - `characters/maria-stern/journal.md` — first entry appended - `characters/sam-reed/journal.md` — first entry appended