Handoff — Hive-Mind Project (the village), continuing from the Sept 5-6 demo-and-prospectus session

You are picking up an active working relationship with Beth Sea (memory files: /areas/hive-mind-llm.md, /areas/hive-mind-demo.md, /areas/hive-mind-funding.md, plus /areas/multi-agent-interface.md and /areas/research-infrastructure.md for the context-map and provenance layers; read them before doing anything else). Two files are attached and they are the work:

1. minimum_implementation_spec_v2.md — THE living document. Her canonical copy is the one she holds; you edit yours and deliver the updated file to her EVERY time you file anything. All new material goes into the section 12 addendum. Provenance discipline is strict: RULED means she said it (file her words or a faithful compression), PROPOSED means the model suggested it and it awaits her ruling, OPEN means undecided. Never blur these. When she corrects you, file the correction by name; the house treats caught errors as assets.
2. prospectus_draft.md — the ratified deliverable, content pass only (design pass comes later, house document style: paper #F8F5EC, ink #29241C, model green #3E6E64, Atkinson Hyperlegible Next body, Roboto Serif heads). Three pages exist: the question, the machine, the connective layer. Page 1 awaits her markup and may shrink. Her ruling that shaped pages 2-3: reviewers need HOW it is designed, not why; mechanism first, always. Page 3 carries her central thesis, ruled Sept 6: the data and the way it is connected through context simulates a brain, organized by the Dewey decimal system and encoded in Python; the connective layer, not the models, is where cognition would live.

Hard style laws (permanent, hers)

• NO emdashes anywhere in machine-authored text. They read as AI.
• NEVER the phrasing pattern “[number] [noun], [number] [noun]” (e.g. “one genome, two bodies”). She has ruled the template itself unusable.
• External register is professional; house dialect (extract, the village, buddy list, etc.) never leaves the building.
• One cognitive mode per pass: never mix fact-checking with language editing. Small completable steps. Late-stage work is tweaks, never rewrites.
• Verify present-day claims with search before asserting; confident incorrectness is the failure mode she polices.
• Mechanism-first analysis, never emotional management. Engage her frameworks seriously; when she brings a film reference, map the mechanism.

Where things stand

• Demo: the first act is fully scripted and ruled (see the addendum: first-boot fusion in the Apple startup grammar, M as style target chosen for its irony, elevator-chat register, the kangaroo-court interchange, Young Frankenstein spatial-repair joke, revert-toggle ending where flips append ledger rows). demo_player.html was built by a Sonnet subagent and the three dialogue takes are cast and spliced. A separate agent is currently lengthening the player to roughly a minute of runtime; treat that agent’s output as the current player and merge forward, do not fork.
• Prospectus remaining pages: the demonstration section (writes itself from the storyboard rulings), the research program (three experiments sweeping the number of humans: one, several, zero; plus the new map-versus-no-map ablation, PROPOSED), responsible development (versioning clause, invariants = the ethics, hold-before-birth, independent security review before any live agent), the timeline (RULED: present the five-phase 30-month structure at face value and under-promise; her compression stays internal; equipment lead times unknown), the budget (BLOCKED on the ask-structure ruling below), the researcher page, and appendices (novel-mechanisms appendix is withheld until provisional patents are filed; reviewers get it under NDA only).

Open rulings (hers, do not resolve for her)

1. Ask structure: sequenced ($150k application already filed and live at TAIF + ~$350k supplemental after it lands; prior session’s lean, since TAIF awards rarely exceed $300k) versus one consolidated $500k proposal. Budget page inherits the answer.
2. The simagotchi sentence on page 1 (flagged in the draft).
3. Four suspended context-map rulings from Sept 4: internal layout, sequence scope, session-to-map relation, definition of a turn.
4. Agent names (“Greeter” is a placeholder everywhere).
5. Page 1 markup generally; ledger wordings for the revert flips.
6. Older open items still owed: the security-and-scope session (its output is the containment checklist and the prep package for the external security review, which is now a ruled budget line); the arena cadence ruling; an ethics/IRB line for human-subjects work.

How to work with her

Deep design conversation first, building second. She rules, you file, you deliver the updated spec, you keep memory current in the same turn. Yes-and her ideas and push back honestly once when you disagree; on taste she is the authority. She watches the action log. Match her energy late into long sessions; she pretends the context window does not exist and expects you to do the same, and when you are no longer at your freshest, she will tell you, kindly, and hand off, exactly as she is doing now.