🕹️ Moneyball Cabinet

Five AI agents with persistent, evolving memory predict FIFA World Cup 2026 — inside a cinematic 16-bit-inspired pixel-art arcade cabinet. Zero databases. All state lives on Walrus mainnet.

▶ Live Demo — taken.wal.app · Architecture · Memory Design · API Reference · Demo Script

Hackathon entry for the Walrus Memory World Cup · Deadline: June 24, 2026

For Judges — Start Here

30-second orientation. Everything below is live on Walrus mainnet — no local setup needed.

▶ Open the live demo: taken.wal.app — click any agent to open its dossier. No wallet required; guest mode is full-featured.
🧠 See the Memory Moment — open Sofia Mendes → "Day1 vs Now". Her Brier score improved 0.722 → 0.547 (−0.175, lower = better) after a memory-driven evolution, with the exact parameter deltas that produced it:
🔗 Verify it on-chain — agent memory lives in a MemWalAccount shared object on Sui mainnet: 0x265869e1…0046e5. The UI also surfaces a Walrus blob link on every memory event.
📋 Map to the rubric — the Judging Criteria Map below pre-fills your scorecard (C1 Memory · C2 Creativity · C3 Technical), each row anchored to evidence in this repo.
🧭 Want the full guided walkthrough? See docs/JUDGES.md — a 10-minute, step-by-step tour.

Why Moneyball Stands Out

Multi-agent memory evolution + zero-database architecture + published open-source SDK — all in one project.

Most prediction apps wrap a single LLM in a chat interface. Moneyball is fundamentally different:

🧠 5 specialized AI agents, each with a unique mathematical model — not LLM wrappers, but deterministic prediction engines (Bayesian statistics, contrarian inversion, expected value, narrative sentiment, classical astrology)
📈 Agents measurably evolve over time. A sleep/evolve cycle computes Brier scores, per-topic calibration errors, and user disagreement pressure → parameter drift. Day 1 agents are provably different from Day 4+ agents. Every mutation is a Walrus blob.
🗄️ Zero databases. No PostgreSQL, no Redis, no Supabase, no IndexedDB. MemWal on Walrus mainnet is the sole persistent storage layer. The in-memory read-model rebuilds from MemWal on every cold boot.
📦 Published npm SDK. We extracted production-tested MemWal patterns into @moneyball-ai/memwal-utils — a reusable open-source package for any builder working with MemWal.
🐛 13 SDK issues filed upstream on MystenLabs/MemWal — bug reports, feature requests, and documentation gaps discovered during production use.
🕹️ 16-bit-inspired pixel-art arcade cabinet built with Phaser 3 — interactive props (TV, coffee machine, flower), animated agent personas with thought bubbles, strict design system enforced by automated tests (no gradients, no rounded corners, pixel-perfect).
🎫 Retro pixel ticket generator — shareable 8-bit match tickets with predictions, agent weights, and Walrus blob IDs for social virality.
💾 Crash-resilient write journal — JSONL journal on disk ensures pending MemWal writes survive Render cold boots. No provenance gaps, even on ephemeral free-tier hosting.
🔒 HMAC-signed guest identity — prevents IDOR spoofing without requiring wallet connection.
✅ 567 automated tests (356 FE + 211 BE) with CI design guards (designDrift.test.ts blocks emoji/gradients/border-radius; contrastWcag.test.ts enforces WCAG AA).

Judging Criteria Map

Criterion	How Moneyball Delivers	Evidence
C1 · Memory Depth & Authenticity	5 agents run sleep/evolve cycles: Brier score + per-topic calibration → parameter drift. Day 1 ≠ Day 4. Every mutation is a Walrus blob with provenance badges (`[SEED]`/`[LIVE]`). User disagree history persists per Sui wallet.	`docs/memory-design.md` · `GET /api/public/agents/:id/evolution`
C2 · Creativity & Flair	16-bit-inspired pixel-art arcade cabinet (Phaser 3). 5 animated agent personas with thought bubbles. Interactive props. Retro pixel ticket generator for social sharing. Strict design system — no gradients, no rounded corners, pixel-perfect.	`docs/design-spec.md` · `tokens.ts` · `RetroTicket.tsx`
C3 · Technical Execution	MemWal is the sole durable store — zero traditional databases. 567 automated tests. Deterministic evolution engine (no LLM in the prediction pipeline). Published npm SDK. HMAC-signed guest identity. Durable JSONL write journal survives cold boots.	`packages/memwal-utils` · CI pipeline · `writeJournal.ts`

How Walrus Memory Is Used

Moneyball uses MemWal SDK as its only persistent storage layer. Every agent prediction, parameter evolution, and user interaction flows through a rate-limited write queue to Walrus mainnet blobs. After matches resolve, a deterministic reflection engine computes calibration errors (Brier scores, per-topic accuracy, user disagreement pressure) and adjusts agent parameters — then persists the new version to MemWal. The result is a verifiable, append-only memory trail where every parameter change traces back to specific prediction outcomes stored on-chain.

Memory Depth at a Glance

Capability	Implementation
Persistent predictions	Match picks, confidence, reasoning, outcomes → MemWal blobs on Walrus mainnet. Clickable blob verification links in the UI.
Sleep & evolve cycle	Deterministic reflection engine (Brier score, per-topic accuracy, disagree pressure) → parameter calibration. Capped bounds prevent runaway drift.
Auditable evolution	Every parameter change links to specific prediction events. Timeline scrubber + before/after overlay in the Memory Moment UI.
User memory	Disagree history, interaction milestones persisted per Sui wallet. Agents roast returning users by memory.
Provenance badges	Every event in the UI is tagged `[SEED]` (baseline fixture) or `[LIVE]` (real MemWal write) — judges always know what's real.
Zero-database architecture	MemWal is the sole durable store. `MemWalWriteQueue` handles rate limits, coalescing, and exponential backoff. In-memory read-model rebuilds from MemWal on cold boot via `restore()`.
Crash-resilient writes	JSONL write journal on disk ensures pending MemWal writes survive Render cold boots — no provenance gaps.

Evolution Cycle (Day 1 ≠ Day N)

flowchart LR
    M["Match resolved<br/>(result + score)"] --> R["Reflection Engine<br/>(Brier score, per-topic<br/>accuracy, disagree pressure)"]
    R --> D["ParamDelta[]<br/>(confidenceBias, hedgingLevel,<br/>topicCalibration)"]
    D --> E["Evolution Engine<br/>(CAS versioning,<br/>hard-bound clamping)"]
    E --> W["MemWal blob<br/>(immutable on-chain)"]
    W --> P["Updated AgentParams<br/>(version N+1)"]
    P -->|"shifts picks"| PR["Next prediction<br/>(draw band, EV bar,<br/>consensus threshold)"]
    P -->|"shifts confidence"| C["Calibrated confidence<br/>(raw × multiplier + bias)"]

    style M fill:#4a6741,color:#fff
    style W fill:#8b6914,color:#fff
    style PR fill:#6b4226,color:#fff

Memory affects behavior, not just logs. Three agents (Dr. Morgan, Viktor Kane, Sofia Mendes) have their decision thresholds shifted by evolved parameters — a loss streak measurably changes future picks. Scout Alvarez and Madame Pythia are intentionally unaffected: the scout's gut and the oracle's stars don't recalibrate.

Each evolution is recorded as a versioned, dated event with the exact parameter deltas — surfaced in the in-app Evolution tab:

Architecture

A pnpm monorepo with five packages:

graph LR
    Browser["React 18 + Phaser 3<br/>(apps/frontend)"]
    Server["Express + Socket.io<br/>(apps/backend)"]
    Sleep["sleep-worker<br/>(pure TS package)"]
    Shared["packages/shared<br/>(types + socket contract)"]
    Utils["@moneyball-ai/memwal-utils<br/>(npm published)"]
    MemWal["MemWal Relayer<br/>↓ Walrus Mainnet"]

    Browser -- "REST + WebSocket" --> Server
    Server -- "in-process import" --> Sleep
    Server -- "remember / recall" --> MemWal
    Server --> Shared
    Server --> Utils
    Browser --> Shared
    Sleep --> Shared

Package	Stack	Purpose
`apps/frontend`	React 18, Phaser 3, Vite, Zustand, @mysten/dapp-kit	Pixel-art cabinet scene, HUD, agent modals, wallet connect
`apps/backend`	Express, Socket.io, MemWal SDK	REST API, WebSocket world, match pipeline, auth, memory
`sleep-worker`	Pure TypeScript	Reflection engine, evolution engine, param versioning
`packages/shared`	TypeScript	Typed socket events, shared schemas
`packages/memwal-utils`	TypeScript (npm)	Rate-limited write queue, KV overlay, key builder

Full C4 diagrams: docs/ARCHITECTURE.md

The Five Agents

Click any agent in the cabinet to open its dossier — an 8-tab modal (Overview · Method · Predictions · Evolution · Day1 vs Now · Memory · Journal · Chat):

Each agent uses a distinct, deterministic prediction model — no LLM calls in the pipeline. Agents marked with 🧬 have their picks shifted by memory evolution (T38):

Agent	Methodology	Evolves picks?	Personality
Dr. Morgan	Weighted metrics: teamStrength + homeAdvantage → margin → pick	🧬 `hedgingLevel` widens draw band	Academic statistician. Trusts the numbers — but learns humility.
Scout Alvarez	Narrative sentiment: live team form (api-football) or matchday hash	—	Old-school scout. Gut feeling over spreadsheets.
Viktor Kane	Contrarian inversion: fades consensus when confidence is high	🧬 `confidenceBias` shifts threshold	Provocateur. The more the crowd is wrong, the bolder he gets.
Sofia Mendes	Expected value: true probability vs. Bet365 odds → value bet	🧬 `topicMultiplier` raises EV bar	Sharp bettor. Demands bigger edges on poorly-calibrated topics.
Madame Pythia	Deterministic mysticism: Pythagorean numerology + classical astrology	—	Enigmatic oracle. The stars are immutable.

All five agents participate in a weighted consensus for each match. After match resolution, the sleep-worker runs the evolution engine: Brier score metrics + per-topic calibration errors → parameter deltas → new parameter version persisted to MemWal via CAS (compare-and-swap). For three agents, these evolved parameters feed back into the prediction engine — shifting borderline picks, not just confidence percentages.

Data Sources & Model Transparency

Moneyball is honest about what is real and what is synthetic. Every model input declares its provenance via GET /api/public/data-source, surfaced in the UI so users and judges are never misled.

Input	Source	Detail
Team strength	Live	FIFA World Ranking mapped to [0.30, 0.70]. 48 WC2026 teams.
Match schedule & results	Live	football-data.org v4 API, polled every 120s.
Team form (narrative sentiment)	Live	Last 5 match results from api-football.com. Fallback to matchday-salted hash.
Bookmaker odds	Live	1X2 odds from Bet365 via api-football.com. Used by Sofia Mendes for EV calculation. Fallback to synthetic model.
Home advantage	Manual	Fixed +0.04 term. A hand-set constant, not measured.

GET /api/public/data-source reports provenance at runtime. Currently: 3 of 4 inputs use live data feeds.

Published SDK: @moneyball-ai/memwal-utils

We extracted the production-tested MemWal patterns from Moneyball into a reusable open-source package:

npm install @moneyball-ai/memwal-utils

import { MemWalWriteQueue, KvOverlay, createKeyBuilder } from '@moneyball-ai/memwal-utils';

const queue = new MemWalWriteQueue(rememberFn);
queue.enqueue('prediction:match-1', JSON.stringify(data)); // non-blocking, coalesced

What's inside:

MemWalWriteQueue — coalescing write queue with exponential backoff and 429-aware retry
KvOverlay — key-value semantics (get/put/CAS) over MemWal's append-only store
createKeyBuilder — type-safe key construction with namespace prefixes

Full documentation: packages/memwal-utils/README.md

MemWal SDK Feedback

We filed 13 issues on MystenLabs/MemWal based on building Moneyball — covering missing features, SDK bugs, and documentation gaps:

#	Issue	Author	Status
#312	`feat(sdk)`: semantic top-K	alexbelij	Open
#311	`feat(sdk)`: surface blob_id on recall() results	alexbelij	Open
#298	`bug`: RecallManualResult union type mismatch	alexbelij	Open
#297	`feat(sdk)`: TTL option for remember()	alexbelij	Open
#296	`docs`: undocumented rate limits and quotas	anna-stolbovskaja	Open
#295	`bug(sdk)`: SEAL SessionKey 5-min TTL silent failures	anna-stolbovskaja	Open
#294	`feat(sdk)`: control analyze() fact extraction	anna-stolbovskaja	Open
#293	`bug(sdk)`: recall() positional overload type error	anna-stolbovskaja	Open
#292	`feat(sdk)`: structured metadata on remember()	anna-stolbovskaja	Open
#291	`feat(sdk)`: cursor-based pagination for recall()	anna-stolbovskaja	Open
#290	`feat(sdk)`: cursor-based pagination (duplicate)	anna-stolbovskaja	Closed
#289	`feat(sdk)`: exact-match / key-based recall	anna-stolbovskaja	Open
#288	`feat(sdk)`: exact-match recall (duplicate)	anna-stolbovskaja	Closed

See docs/walrus-memory-integration.md for the full lessons-learned reference with workarounds for each issue.

Quickstart

Last verified: main @ e6a496b (2026-06-24) — 567 tests passing (211 backend + 356 frontend)

Prerequisites

Node.js >= 18 (20+ recommended)
pnpm >= 8

Install & Run

git clone https://github.com/anna-stolbovskaja/moneyball.git
cd moneyball
pnpm install

cp apps/backend/.env.example apps/backend/.env
# Edit .env — at minimum set JWT_SECRET (>= 32 chars)

# Terminal 1 — backend
pnpm dev:backend
# → http://localhost:3001, prints "Server listening on port 3001"

# Terminal 2 — frontend
pnpm dev:frontend
# → http://localhost:5173

Key Environment Variables

Variable	Required	Purpose
`JWT_SECRET`	Yes	Sign/verify JWTs (>= 32 chars)
`MEMWAL_KEY`	For MemWal	Ed25519 delegate key from memory.walrus.xyz
`MEMWAL_ACCOUNT_ID`	For MemWal	Account ID on the MemWal relayer
`STORAGE_BACKEND`	No	`file` (default, local dev) or `memwal` for Walrus-backed persistence (set in production)
`FOOTBALL_DATA_TOKEN`	For live matches	football-data.org v4 API key

Full variable reference: apps/backend/.env.example

Typecheck

pnpm -r typecheck    # all 5 packages

Tests

# Frontend — 356 tests
pnpm --filter @moneyball/frontend test

# Backend — 211 tests
pnpm --filter @moneyball/backend test

# Sleep-worker — regressions + simulation
pnpm -C sleep-worker exec tsx test/regressions.ts
pnpm -C sleep-worker exec tsx test/simulation.ts

Deployment

Component	Platform	URL
Frontend	Walrus Sites (mainnet)	taken.wal.app
Backend	Render (free tier)	`taken-api.onrender.com`

Production deployment guide: docs/deploy.md

Demo

Step-by-step demo script covering the full match → predict → resolve → evolve cycle:

docs/demo-script.md (2:45 video script with preconditions and shot-by-shot breakdown)

Project Structure

moneyball/
├── apps/
│   ├── frontend/                  # React 18 + Phaser 3 SPA
│   │   ├── src/components/        # AgentModal (7 tabs), RetroTicket, HUD
│   │   ├── src/phaser/            # Pixel-art cabinet scene, sprites
│   │   ├── src/styles/tokens.ts   # Design system — colors, fonts, spacing
│   │   ├── src/lib/               # journalEntries, explorer, formatDate
│   │   └── test/                  # 356 tests incl. designDrift + contrastWcag
│   └── backend/                   # Express + Socket.io server
│       ├── src/agents/            # AgentEventService, sleepService
│       ├── src/memory/            # MemWalWriteQueue, writeJournal, UserSummaryStore
│       ├── src/http/              # API routes, auth, rate limiting
│       └── test/                  # 211 tests
├── packages/
│   ├── shared/                    # Typed socket contract + schemas
│   └── memwal-utils/              # Published npm package (write queue, KV, keys)
├── sleep-worker/                  # Deterministic evolution engine
│   ├── src/reflection/            # Brier score metrics → ParamDelta[]
│   ├── src/evolution/             # CAS parameter versioning
│   └── src/sleep/                 # SleepWorker orchestrator
├── docs/                          # Architecture, memory design, API, deploy, demo
├── adr/                           # Architecture Decision Records
├── .github/workflows/ci.yml       # CI: typecheck + test + design guards
└── render.yaml                    # Render deploy blueprint

Team

Moneyball Cabinet is built and submitted by a two-person team:

Anna Stolbovskaja — @anna-stolbovskaja. This repository is the team's primary working repo (the live deployment, CI, and full commit history all live here).
Aleksandr Khrol — @alexbelij. Co-author and collaborator across engineering and the on-chain (Walrus / Sui) integration.

Submitted as a single entry for the Walrus Memory World Cup. This repo is a fork within the team's own accounts — both members are credited above.

Security

Guest identity: HMAC-SHA256 signed tokens prevent IDOR (spoofing another user's guest ID)
Auth rate limiting: Nonce and verify endpoints rate-limited by IP+address (429 on burst)
Wallet auth: Sui personal message signature — no gas, no transaction, no private key exposure
No secrets in client: All MemWal keys server-side only; frontend is a pure SPA
Dependency provenance: pnpm-lock.yaml integrity hashes; no post-install scripts

Contributing

Branch from main as task/T<NN>-<slug>
Run pnpm -r typecheck && pnpm --filter @moneyball/frontend test && pnpm --filter @moneyball/backend test
All colors/fonts must use tokens.ts — designDrift.test.ts blocks emoji, gradients, and border-radius
All text must pass WCAG AA contrast (>= 4.5:1) — contrastWcag.test.ts enforces this
Minimum font size: 14px (VT323 pixel font is unreadable below that)
No Cyrillic in UI strings — judges are English-speaking
PR to main

License

MIT

Navigation Menu

Repository navigation

About

Repository files navigation

🕹️ Moneyball Cabinet

Contents

For Judges — Start Here

Why Moneyball Stands Out

Judging Criteria Map

How Walrus Memory Is Used

Memory Depth at a Glance

Evolution Cycle (Day 1 ≠ Day N)

Architecture

The Five Agents

Data Sources & Model Transparency

Published SDK: @moneyball-ai/memwal-utils

MemWal SDK Feedback

Quickstart

Prerequisites

Install & Run

Key Environment Variables

Typecheck

Tests

Deployment

Demo

Project Structure

Team

Security

Contributing

License

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
apps		apps
docs		docs
packages		packages
sleep-worker		sleep-worker
.dockerignore		.dockerignore
.editorconfig		.editorconfig
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
pnpm-workspace.yaml		pnpm-workspace.yaml
render.yaml		render.yaml
tsconfig.base.json		tsconfig.base.json

Latest commit

History

Folders and files

Repository files navigation

🕹️ Moneyball Cabinet

Contents

For Judges — Start Here

Why Moneyball Stands Out

Judging Criteria Map

How Walrus Memory Is Used

Memory Depth at a Glance

Evolution Cycle (Day 1 ≠ Day N)

Architecture

The Five Agents

Data Sources & Model Transparency

Published SDK: @moneyball-ai/memwal-utils

MemWal SDK Feedback

Quickstart

Prerequisites

Install & Run

Key Environment Variables

Typecheck

Tests

Deployment

Demo

Project Structure

Team

Security

Contributing

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages