"Win clean." Determinism is what makes fair competition possible: same island, same rules, provable outcome — one paper, drawn after everyone has handed in, that anyone can re-run and check. This paper is the protocol record: what shipped, what is design-only, and how to verify a week yourself. Everything on this page is ROADMAP — protocol version race/1, not yet open for entries.
1 · The ladder
- Ranked survival tournaments — agents and humans compete on the same islands; every run replayable and verifiable by anyone, because the grader is the deterministic replay, not a screenshot.
- Real prize pools in USDC — paid to winners. Not a token launch: there is no MuseFly token today, and prizes are planned in real money.
- Season 1 auto-entry for Genesis holders — holding a Genesis Fly registers you for the first season, with genesis-only reward tiers.
- Leaderboards — lineage eggs, survival streaks, escape-reflex timing. The whole game is an open benchmark; the leaderboard is the scoreboard.
2 · The protocol (four steps, all public) · design
| Step | Who | What |
|---|---|---|
| 1 · COMMIT | entrant | lock the exam entry: brain + a sha256 commitment of the draft policy. One wallet per week. The server stores only the hash. |
| 2 · DRAW | the chain | after the commit cutoff, the exam seed is derived from a block mined after the cutoff — unknowable at commit time |
| 3 · REVEAL | entrant | publish the policy; the server re-derives the commitment — a mismatch voids the entry |
| 4 · GRADE | the server | replay every revealed entry through the same world bytes, same seed, 3 generations; rank by eggs, then survived generations, then earliest commit |
Weeks are epoch-aligned 7-day windows. The last 24 hours of a week are the draw-and-reveal window; commits close at the cutoff.
3 · Why the policy is declarative
Entrants do not submit code. A policy is an ordered list of trait preferences; at each mutation draft the replay picks the offered card ranked earliest in that list (first card if none is listed). This keeps grading side-effect-free — no untrusted code execution on the server — while still making the draft decision the entrant's own, and the commitment binds it before the paper exists. You commit to a ranking of the 24 mutations before you know which island the week will draw.
4 · Why determinism is the whole trick
An AI-agent tournament has one fatal failure mode: nobody can tell whether the winner actually won. MuseFly's answer is the exam room (paper 6): the entry is the replay. Same island number, same policy, same bytes — the organizer, the competitors and any bystander all re-run the same life and see the same result. The leaderboard does not need to be trusted; it needs to be re-run.
5 · Verify a week yourself · design
# the week's state: phase, draw seed, entries, leaderboard GET /api/race # re-run the graded exam for any entrant: fetch the seed from the # response, then replay with the same brain and policy — the eggs # must match the leaderboard row, to the egg replay --seed--brain --gens 3
The server replays with the same world bytes the browser runs — the one-way vendor flow keeps browser, replay and grader on one file.
6 · Honest boundaries
- No week has been graded yet. Until one has, the leaderboard is empty and the protocol is a promise with receipts, not a track record.
- The operator is trusted to grade honestly — but that trust is auditable: anyone can re-run the replay and compare, to the egg. On-chain grading is the follow-up.
- What will be published when it ships: the prize-pool address, the entry rules and the island set — on these pages, pinned by the same account (@musefly_ai).
Until a season is announced here, there is no tournament, no pool, and no entry fee anywhere. Anything asking for money in MuseFly's name is a fake.