discipline

analyst-desk copy rather than a live conversation.

## Vocabulary discipline

A few rules-of-the-road that keep the language tight across docs and prompts:

1. **Prefer the canonical term over a paraphrase.** "Priors" not "accumulated knowledge" or "queue" or "between-session input"; "readings" not "scenario picker".
2. **System / World / Fate are always capitalised when naming the forces** — even mid-sentence — to distinguish from the everyday senses of those words. "the System force" / "the domain's *world* is rich" (lowercase only when it's clearly the noun, not the force).
3. **"Domain" and "domain" are interchangeable.** Domain is the canonical case study; domain is the underlying abstraction. Use whichever fits register. Avoid "model" or "simulation" as substitutes — both have other meanings here.
4. **PRG = Phase Graph** in the UI; `mode` survives only as a compatibility identifier. The per-Arc CRG is **Causal Graph** in compact interface copy and **Causal Reasoning Graph** on first educational use.
5. **The room is the application; substrate is the engine.** Never call the substrate "the room" or "the engine" interchangeably. Lead with the writers'-room / table-read / playable-world framing; "War Room" survives only as the legacy alias for the strategic / non-fiction mode.
6. **Private and Public are first-class adjectives, not modes.** "A private room" / "a public game on a hosted substrate". Not "private mode".
7. **Cards are intent signals.** Avoid "moves" (the substrate uses "move" loosely for the underlying action); reserve "card" for what the player actually plays.
8. **Stakes are a layer, not a feature.** Stakes are *attached to plays*, optional, opt-in. Avoid calling them "the betting feature" or "the gambling layer".
9. **Don't introduce new vocabulary without adding it here.** If a recurring concept shows up in two or more docs and isn't in LANGUAGE.md, it either belongs here or it's drift that should be corrected back to an existing term.
10. **Console / live (ngrok tunnel) / dark (Signal) is the access language.** The console is the master device (the Director's computer that holds the substrate and drives the shared screen), running an **always-on Electron host** that keeps the instance and tunnel up continuously — the main way a room stays on. **Live**, non-Directors reach the *full interface* over a PIN-gated **ngrok tunnel** — a **stable, always-on** public HTTPS URL (read-write; Director elevated; app-layer auth is the perimeter; Director-revocable; LAN-only fallback for fully-private tables). **Dark**, the async capture layer runs over **Signal** — the one capture channel, **end-to-end encrypted** (dedicated number, DMs not a group, quarantined ingress); the Director curates and commits. There is **no central server** (the tunnel is third-party; Signal = a normal chat; the Director machine is the always-on host). In user-facing copy don't call the console "the server"; the tunnel view is the **real interface**, not a "crippled mobile view." *(ngrok for the always-on stable URL; Signal for sovereign/team dark capture; a **Telegram bot** is the separate **consumer-rung** surface — a trained, evolving chatbot the Director schedules to brief them (notification / query / control over the artifacts, where the compounding intelligence lives behind the chat), the B2C growth-engine channel kept distinct from team capture; Tailscale not used.)*
11. **Priors are human-made, never scraped.** The prior-collection is humanistic by design — a person filters, structures, and sets the threads; web search only *assists* drafting. Don't reintroduce "automated feeds" or "scrapers" as inputs.
12. **Two map types only: Graph and Board.** The raw graph substrate and a board-game-style map with nested maps. Don't reintroduce "grid" or "hex" as board geometries.
13. **The queue is now "Priors."** "Queue" / "Driver" survive only as code/UI labels inside the **Signals** (Capture) surface; in prose and vocabulary it's *Priors*.
14. **Player-archetype tag `id` matches its `label`** (kebab-case) in the game-theory dashboard — no literary/code-name drift between what's stored and what's shown. New tags get an entry under *Player archetypes* above.
15. **A decision is 1- or 2-player.** Two-player = *duel* (a matrix); one-player = *solo* (a row, reality in the other seat). Don't describe game theory as inherently two-player.
16. **Threads, Streams, and Positions are distinguished by the WRITE PATH, not by outcome-vs-action.** All three are **predictive reads over mutually-exclusive outcomes, moved by evidence**; what separates them is where the read goes. A **Thread** is the domain's own canonical belief — *canon*. A **Stream** is one perspective's seat-owned, evidence-accruing read that **stages a change to canon** and terminates in a **Merge**. A **Position** is a composed forward read that **never writes to canon**. **An action is an outcome with a known driver** — when a Stream's question is one the seat itself decides, its outcomes *are* its candidate moves; there is no separate kind, field, or code path. Don't use "thread" for a seat's per-perspective read, or "stream" for the canonical domain belief. *Retired: "**Streams are action-first**" — action is one framing a Stream's question may take, never its identity, and in a Public Domain it normally does not.*
17. **The loop is *Model → Capture → Scenario*.** Three beats, no fourth. *Review* and *Butterfly* are retired as loop phases — don't reintroduce them; the slide-deck output lives on as **Onboarding decks**.
18. **Scenario (capital) is the live practice / game; conviction (lowercase) is the resource.** Scenario is the live practice (was "Playback"), played as the **Scenario** card game — scenario is a pluggable layer, so other live formats can ride the same substrate later. *Scenario* names the act (playing the story forward); *Scenario* names the practice it's played as. Don't call conviction "the betting chips".
19. **The Learning layer is additive and post-hoc.** Question banks, scoped practice, Curriculum, and Coverage read the domain; they never mutate force deltas. Don't frame Learning as part of generation.
20. **Capture is the asynchronous practice; Scenario is the live one.** Tempo is the clean split — Capture runs *async* (priors recorded over time, between sessions, no convening), Scenario runs *live* (a synchronous, high-feedback session the room plays together, acting the story forward against AI scene partners). The two modes of the Vision thesis (*human judgement and vision are the edge the frontier models don't supply*). Don't blur it: don't call Capture "live" or Scenario "asynchronous" at the practice level; "Playback" is retired in favour of **Scenario**. (The access model's *live (ngrok tunnel) / dark (Signal)* is a separate, device-reach layer — orthogonal to this practice-tempo split.)
21. **Read / Write / Play is the player's loop; Generate is the Director's beat.** Players carry only the three-beat **Read → Write → Play** (poker-inspired, low-friction — the onboarding surface). The fourth beat, **Generate** (the world simulated one step forward), plus all facilitation, is the **Director's** job — handle the details, keep the room social and fun. Don't burden the player-facing loop with the engine machinery; that lightness is what makes Scenario an onboarding experience.
22. **Lead B2B with worldbuilding teams; the creator network is the volume tier beneath.** The first buyer is the team with a real worldbuilding budget — studios, showrunners, dev execs, game/animation worldbuilders. Indie writers and creators are the downstream volume tier, reached bottom-up through the **Player → Contributor → Director** ladder (card game as funnel, directors as distribution) — a *network that grows itself*. Don't describe Meridians as a "single-player forecasting tool," a "consulting practice," or a "services-led boutique." This is the deliberate fix for the old monetization hole: worldbuilding is a real budget line; "ideation" sold to cash-poor indies was not.
23. **`master` / `client` are the only sync-role words — `clone` and `tributary` are retired.** The runtime type is `SyncRole = "master" | "client"`; use those in docs, comments, and copy. "clone" survives *solely* for *duplicating* a domain/scene (a domain clone), never for the joined origin. When you see `clone`/`tributary` for the sync role, it's drift — read it as **client**.
24. **`Director` (platform account) vs `director` (in-instance role); `Admin` (platform operator) vs `manager` (in-instance role).** Two clean splits: capital-**Director** = the Supabase account owner, lowercase **director** = the top ladder role (they're the same person, different systems). **Admin** = the platform operator (`platform_admins` / the gateway `/admin`); **manager** = the ladder role just below director. **There is NO in-instance "admin" and NO platform "manager"** — the words don't cross layers. (The manager role was renamed from "admin" precisely to free "admin" for the platform; don't reintroduce "admin" as a member role, and don't call the platform operator "Owner" — it's the **Admin**.)
25. **master/client ⊥ License/Hosted — two orthogonal axes.** *master/client* names **runtime authority** (who holds the record); *License/Hosted* names **the product** (where the master runs + who pays/keys). A Hosted VM is a **master** on our infra; a License app is a **master** on the user's machine; collaborators are **clients** in both. Don't say "hosted" to mean "client" or "master" to mean "License-only". After E1 the **master is the server-of-record daemon**, not the browser — the browser is the local **client**.

Raw Markdown source