Proof of Work pays for waste.Proof of Frontier pays for progress.

Every day there is a new market puzzle. Everyone plays the same one. The best score anyone has proven is the frontier.

You go long or short on each stock so that your basket follows a target. A zero-knowledge proof shows your score without showing your basket. Proven scores mine . New records share a pot of their own.

The compute people spend goes into a real optimization problem, not into hashes that nobody needs.

The puzzle

Each day has a target stock, for example NVDA. You get a list of other stocks. For each one, pick a side: long or short.

Your basket has a PnL curve: the sum of your longs minus your shorts, session by session. The puzzle is to make that curve follow the target. The space between the two curves is the gap. Your score is the gap, and lower is better.

That is the whole game. Flip a stock, watch the gap. Small flips can help or hurt, and they interact: a flip that helps now can hurt after three other flips.

The formula, for those who want it

Two puzzles a day

Same target, two sizes. The Sprint is free practice. The Frontier is the race that pays.

You buy tickets with ETH. Each Frontier submission burns tickets. Tickets cannot be sent to another wallet, and burned tickets do not come back. Tickets are game credits, not . On the testnet, the game never burns . 85 % of what tickets cost buys for the vault. The team's 15 % is paid in ETH, so the team gets no from the game. On the mainnet, the whole ticket buys : 10 % of it is burned and 90 % goes to the vault. The team takes nothing from tickets.

The frontier & how you earn

The Game has a vault of . Each day it releases 1 % of the free vault for the Frontier pot. The pot has two parts: the floor pays every good score, and record steps pay whoever moves the frontier.

The floor: get under the line

On the mainnet, the floor pot is split in equal parts. Each wallet with a proven score at or under the floor line gets one part of the day, whatever its score under the line, its number of tries or the time it arrived. More tries from the same wallet keep its best score and add no part. Each part is the floor pot divided by the number of wallets in the floor when the day ends.

On the testnet, every proven score under the floor line takes a place in a queue, counted in tickets. The earlier your tickets are in the queue, the more they earn: up to 2× the base rate at the head, then less and less. The floor pot is limited. Late or weak submissions can earn less than the tickets they burned. The floor never sells $FRONTIER below the price the Game pays for it: once the Game has bought $FRONTIER today or yesterday, a floor ticket pays at most 85 % of its cost at that price. Without such a buy, the floor pays nothing.

Record steps: move the frontier

On the testnet, when your score beats the record, you open a step worth ln(old record / your score): every halving of the gap is worth the same, near the floor line or near the best. At the close, the record pot is shared in proportion to steps. A tie with the record opens nothing, so copying the record earns nothing from the record pot. Cutting one improvement into many small steps pays the same total and costs more tickets.

On the mainnet, the prize pot goes to the top 10 best scores of the day: 40, 20, 12, 8, 6, 4, 3, 3, 2 and 2 % of it. Each wallet takes one place, with its best score. On a tie, the earlier submission is listed first, and the tied wallets share the pay of their places equally. With fewer than 10 wallets, the places shown share the whole pot.

Why pay for progress, not for compute

Bitcoin pays in proportion to hashes. Here, compute pays only through results: a score under the floor line, or a better score than anyone has proven. A smarter search can beat a bigger machine, and the reward follows the result, not the effort.

• Records follow steps, not wallets: one improvement split over 100 wallets earns the same. On the mainnet floor, each wallet in the floor gets one equal part, and each new wallet needs its own ticket and its own proof.
• What nobody earns goes back to the vault: the rest of the floor pot, a record pot nobody opened.
• Supply: 1 billion , created once at the launch (on Pons for the testnet, on Abyss for the mainnet). Nobody can create more. 85 % of what tickets cost, and 85 % of the 3 % tax on every trade, buy back for the vault.
• You claim after the day closes, within 30 days. After that, what nobody claimed goes back to the vault.

Set at launchThe testnet releases 1 % of the vault a day. It still keeps 5 % of each day for the old Sprint pot, which goes back to the vault when nobody wins it. On the mainnet, a ticket costs 0.01 ETH, paid in ETH or in at the same value. Paying in works only within one hour of the Game's last buy. Each day pays out what the tickets of the day before put in the vault, plus a match from the vault: 30 % on days 0 to 2, 20 % on days 3 to 6, then 10 %. A day gets at least 6 % of the free vault, and never takes more than 10 % of it from the vault. Day 0 gets 10 % of the free vault. 80 % of a day goes to the top 10 and 20 % to the floor. The Sprint stays free. On the mainnet, each trade pays a 2 % fee. These values are not final until the launch. The owner is a 24 h timelock: every change is public a day before it applies. New reward values and submission costs never alter a day that is already open. The launch values will be published before the mainnet launch.

Why it's a real competition

A leaderboard is only worth something if nobody can freeze it. We attacked this puzzle with the strongest methods we know before we chose it.

• It is hard in general. The puzzle is a QUBO. Finding the best ±1 choice for this kind of problem is NP-hard in the worst case. That is a statement about the worst case, not a promise about each day.
• Strong searches do not agree. Four independent 15-minute searches on the 256 × 32 size ended at scores 87 % apart, with no ties. Nothing converged.
• Specialist attacks fall short. Karmarkar-Karp, a balancing greedy and Lovett-Meka, each followed by local search, ended 1.17× to 3.9× worse than our best basket. Lattice reduction alone (BKZ, blocks 20 to 40) ended about 10× worse on a 128-stock test. A constraint solver that started from our best basket found nothing better in 5 minutes.
• Copying is useless. The target and the stocks change every day. Yesterday's basket does not solve today's puzzle.

Our honest limit: each of these tests ran for 1 to 30 minutes on 4 threads, not for weeks on GPUs. The best possible gap is not known. That is exactly why the record can keep falling.

Why it matters

Under the game is a textbook problem: choose +1 or −1 for each of 256 series so that their sum follows a target. Other fields need the same answer.

The format used to test quantum machines and solvers

The puzzle is a QUBO, the standard benchmark format for Ising machines and quantum annealers (Mohseni et al., Nature Reviews Physics, 2022). No single heuristic wins on every instance, and most studies use artificial instances (Dunning et al., INFORMS JoC, 2018). Ours come from real market data.

5G and 6G receivers

Finding the ±1 values that best explain a signal is optimal MIMO detection, and it is NP-hard (Verdú, 1989). Researchers run it on quantum annealers (Kim, Venturelli & Jamieson, ACM SIGCOMM, 2019). One difference: our puzzle has more unknowns than measurements (256 for 32).

Index tracking

Following an index with a limited set of stocks is a hard optimization problem (Beasley, Meade & Chang, EJOR, 2003). Our puzzle is a simple form of it: long or short, equal weights. It is the same optimization, not a trading strategy: we measured that a record basket does not follow the target outside its 32 sessions. It is a benchmark, not a basket to buy. US index mutual funds and index ETFs held $16.2 trillion at the end of 2024 (ICI Fact Book, 2025).

A fair test for AI search

Record-setting AI systems need a problem with an automatic judge (FunSearch, Nature, 2024; AlphaEvolve, 2025). This puzzle has one. A fresh instance every day cannot be in anyone's training data, the idea behind live benchmarks (LiveBench, ICLR 2025).

What the world getsA new instance every day, and public, proven records on it. Each record proves that a basket with that score exists. The baskets stay secret.

What we do not claim

• Not the best possible basket. "Record" means the best score proven so far.
• Not a portfolio. Outside its 32 sessions, a record basket does no better than a random one (measured).
• Not that quantum computers play here or beat anyone. This is the format they are tested on.
• Not that FunSearch or AlphaEvolve were tested on this puzzle.
• Not training data for AI. Without published baskets, it is a benchmark.
• Not the first useful proof of work. Proof-of-Search (2019) came before. What is new is the mix: a daily game, proven scores, public records.

Fair by design

Nobody picks the puzzle, nobody sees your basket, and nobody can take your score.

1. A public seeddrand
   Each day starts from a drand round, a public random value that nobody controls. The chain checks its signature before the day opens.
2. A public puzzlereplayable
   The seed picks the target and the stocks. The closing prices are public. The puzzle's fingerprint goes on the chain, and anyone can rebuild the puzzle and compare. A mismatch would be public proof of cheating.
3. Your basketprivate
   You build it in your browser. It never leaves your device.
4. Your proofzero-knowledge
   Your device makes a proof that your basket has your score, on this puzzle, for your address. It shows the score and nothing else. In our tests, a Frontier proof took about 5 seconds on 4 threads. A phone can take longer.
5. The chain checks it
   A wrong score, a wrong puzzle or someone else's address: the proof fails. A copied proof is useless, because it is bound to the wallet that made it.

Mine in your browser

The app has a Mine button. Press it and your spare CPU cores search the Frontier puzzle while the tab is open. Your best score updates live. You stop it with one click. It never submits by itself: each submission burns tickets, so you choose when.

Scripts and bots can play too. The proof checks the score, not who or what found it.

Play with your AI agent

Your agent can play for you. Copy this message into Claude Code, Codex or any agent that can run commands on your computer.

• What it does. It downloads a small player from this site, finds the best basket for today's puzzles and builds the proof on your computer. Your basket never leaves it. The next day, it can collect what you won.
• What it needs. Node 20 or newer. To send a score: a wallet with a little ETH on Robinhood Chain, which your agent reads from your environment. Never paste your private key into a chat.
• What it can cost. Nothing until you say yes. Each submission burns tickets, like a submission in the app.

The full guide for agents: agents.txt (short) and AGENTS.md (complete).

FAQ