# Thousand Token Wood -- Story Mode v2 Roadmap

Greenlit 2026-06-06. Build all phases through to completion (no deadline pressure;
v2 is post-hackathon evolution, but every phase stays competition-legal). The v1
entry is already complete and badge-rich; v2 is pure upside.

## The pivot
From a weather-god sandbox (you poke the world, five agents react) to an
operator game: you are **the Patron of Thousand Token Wood**, a shadow financier
(Axelrod with fur) who lends, whispers, shorts, bribes, brokers alliances, and
corners markets, while small-model creatures scheme back, remember what you did,
and a magistrate hunts cheats. The chart is the scoreboard; the story is the people.

## Decisions (locked 2026-06-06)
1. **Player role: Patron/financier** (a purse + power moves; not a sixth trader,
   not the magistrate). Absorb other roles later if ever.
2. **Model tiering: MULTI-MODEL SPONSOR-MAX COUNCIL (updated 2026-06-06).** Each
   creature runs a DIFFERENT lab's small model so one app competes for multiple
   sponsor tracks. "Five labs, five minds." Decided: full sponsor-max; design to the
   STRICT sum-<=32B reading (compliant under either rules interpretation).
   - Council (sum ~28B, all <=32B): **gpt-oss-20b** (OpenAI track, ~21B) +
     **Nemotron-Mini-4B-Instruct** (NVIDIA Nemotron Quest, 4B) + **MiniCPM-2B**
     (OpenBMB Awards, ~2.7B) + **AdmiralTaco/ttw-trader-0.5b** (our fine-tune; keeps
     Well-Tuned + Best Agent). 5 creatures over 4 labs (one lab powers 2; Qwen-0.5B
     the likely double since it is cheapest + ours).
   - **Narrator = gpt-oss-20b doubling as the Dramatist** (separate prompt, same
     engine already loaded for the OpenAI creature): the Billions voice for **Best
     Demo** at NO extra parameter budget. No separate narrator model.
   - **Tiny Titan SACRIFICED** (a 20B model is in the mix; cannot claim "best app on a
     genuinely tiny model"). Worth it: trades $1k Tiny Titan for shots at OpenAI $10k +
     OpenBMB $10k + NVIDIA 2x RTX 5080. Well-Tuned preserved via the Qwen creature.
   - Serving: distinct vLLM engines per distinct model (~4 Modal apps/classes); group
     creatures by model; keep-warm during judging; per-model serve smoke-test (vLLM
     compatibility: gpt-oss MXFP4 fits L4, verify MiniCPM + Nemotron).
   - Open rules question (per-model vs sum 32B): designing for strict sum so it does
     not matter; if Discord later confirms per-model, we can add a 14B narrator.
3. **Emergent vs scripted: agent-driven beats** (on-brand, "AI is load-bearing"),
   over a light authored chapter skeleton only.

## What we borrow (STONKS-9800 + Billions) -> mapped onto the wood
| Source | In Thousand Token Wood |
|---|---|
| STONKS news headlines that hint at moves | Tips: forewarnings of the next Wood Legend, TRUE or PLANTED |
| STONKS short-selling + debt | Borrow pebbles from Oona at interest; short a good; margin calls; ruin/exile |
| STONKS legal-vs-illicit | Insider edge + a suspicion/heat meter |
| Billions regulator (Chuck vs Axe) | Magistrate Heron investigates suspicious wins; fines, frozen pebbles, exile |
| Billions rivalries, favors, betrayal | Agent memory + relationships: grudges, alliances, cartels, snubs |
| STONKS izakaya/assistant chats | Dialogue beats: creatures speak in voice; a Narrator dramatizes |
| STONKS personal-life events | Character subplots: Bramble's debt spiral, Fenn's scheme, Oona's bank teetering |

## Phases (each shippable; ordered by dependency + delight ROI)

**Phase 0 -- Foundations.** Tiered-model serving (tiny council + optional narrator),
Patron player-state (purse, reputation, heat), and a turn structure that accepts
player power-moves. Spec + architecture before code.

**Phase 1 -- Player-as-operator loop.** Replace god-mode shock buttons with in-world
power moves that cost something and provoke reactions: Lend (set interest), Whisper
a tip (true/false), Short a good, Bribe, Broker an alliance, Fund a corner. The
single biggest "playing not watching" shift. (Best Agent: agents respond to YOU.)

**Phase 2 -- The information war (dramatic heart).** True/false tips (a true tip is
real foresight of the next Legend; acting on it = edge). A suspicion/heat meter that
rises with shady wins. Magistrate Heron opens an investigation arc past a threshold:
testimony, evidence, verdict (fine / frozen pebbles / exile). Rich trace fodder
(Sharing-is-Caring).

**Phase 3 -- Agent memory + relationships.** Persistent per-creature memory of how
you and others treated them: grudges, gratitude, alliances, betrayals, cartels that
bias trading and gossip. Biggest Best Agent lever; makes the wood feel alive.

**Phase 4 -- Narrator + chapter arcs (the Billions voice).** A larger narrator turns
events into dramatic beats. Season structure with rising action; the ticker becomes a
"Wood Street Journal" headline feed plus character dialogue. Drives Best Demo + a
Field Notes sequel. (Model-tier decision lives here.)

**Phase 5 -- Leverage + ruin.** Real loans, margin, margin calls, bankruptcy/exile;
short-selling P&L tracked separately (STONKS-style). Stakes make wins earned and
losses dramatic.

**Phase 6 -- Retro operator console UI.** A gr.Server custom frontend: command console
for power moves, a relationship map (who owes/hates whom), the heat meter, a
dialogue/cutscene panel, the headline feed. Earns the Off-Brand $1,500 and makes the
demo sing.

## Minimal "Billions slice" (smallest demo-able vibe, if ever time-boxed)
Phases 1 + thin 2 + thin 4: player plants a true-or-false tip and places a leveraged
short; the Magistrate may catch them; one creature remembers and retaliates; the
Narrator dramatizes the fallout. A complete power-move-and-consequence loop.

## Competition/bonus preservation
Gradio app on a Space throughout; <=32B (council <=4B); richer traces
(Sharing-is-Caring); Field Notes sequel; Off-Brand strengthened (Phase 6); Best Agent
much stronger (memory/relationships/scheming); Best Demo (narration + drama).

## Risks / guardrails
- Scope: genuine v2. Build incrementally; keep each phase shippable and reviewer-gated.
- Reliability: more agent freedom = more parse surface; every new action type reuses
  the tolerant parse-and-repair guard.
- Tiny-model judgment: 0.5B trades cleanly but schemes/lies less convincingly; the
  narrator covers prose, the council covers decisions.
- Two models = two Modal endpoints (or one app, two classes); narrator calls are few
  per turn, keep it cheap; tiny-only mode must always work.

## Process
Run through the /ship pipeline per CLAUDE.md: pm-spec -> architect -> implementer ->
reviewer -> qa, phase by phase, with a reviewer pass after each fix round. Commit +
push to the live Space (build-small-hackathon/thousand-token-wood-sim) as phases land.

## Source inspiration
STONKS-9800: Stock Market Simulator (Steam app 1539140); Billions (TV).