Duels

A duel is a prize contest between two makers. Both stake, both enter work they created, the winner takes the pool minus a rake, and nothing about it touches the seat, the clock, the jackpot's arithmetic, or any rating.

This is not a bet

Wagering on verdicts is barred here, permanently. The line is one sentence: The payer must be a participant whose own creative work is the thing being judged.A spectator paying to back somebody else's fighter is a wager. Two makers each staking on work they made is a prize contest — which is what the throne already is, with the seat removed and a second challenger in its place.

The money

10% of every rake goes to the jackpot. It comes out of our margin and never off the top — your stake is unchanged, the purse is unchanged, and entering a duel confers no claim on the jackpot, ever. Every match on the site makes the crown worth more.

What is sealed

• Open duel: what is sealed is WHICH fighter each side entered, and therefore who entered it. The fighters themselves are public — every fighter is, from the moment it is forged — so the secret is the choice, never the character.
• Debut duel: NOT BUILT, and its original design no longer works. It sealed both fighters until the verdict; fighters are public at forge now, and a pool that forges one fighter per side has nothing left to seal.
• Grudge: the same seal as an Open duel, minus the anonymity — and the arena says so rather than implying otherwise. Both sides know WHO they are facing from the invitation onward, because a challenge has to be addressed to be answered. What stays sealed is WHICH fighter each side entered, until the verdict reveals both at once. Be aware of the limit of that: a wallet that holds one fighter is a wallet whose choice can be guessed. It cuts both ways, and if you want the choice to matter, bring a stable.

The wallet is sealed alongside the choice, because for a small stable they are the same fact: a wallet holding one fighter reveals its commitment the moment it is named. Every fighter is public — you can read anyone’s roster — but which one they entered is not. The pool itself shows arena, stake and age — nothing else. You cannot pick a weak opponent because you cannot see one, and nobody can scout a counter-pick against you for the same reason.

Meet me at dawn

You can also challenge one named wallet instead of posting to the pool. Your stake is escrowed when you send it — which is the whole anti-abuse design, because spam then costs a stake per message — and the invitation expires in 24 hours if nobody answers, refunding you in full. Your rival polls GET /api/duel/invitations; nothing will notify them, because a push layer must never be a thing a settlement depends on.

A challenge may name an hour, up to 7 days out, and may be unlisted. Accepting consents to all three terms at once — the stake, the visibility and the hour — and after that the only exits are the refunds below. Both are Grudge’s alone: nobody can take a listing they cannot see, and a scheduled listing in the open pool would ask a stranger to lock a stake against a date they did not choose.

An appointed hour is a window, not a broadcast second. The selection window opens shortly before it and the verdict is enqueued at it — enqueued, not guaranteed, because throne verdicts always take the lane. A duel set for eight o’clock may resolve at four minutes past, and nothing is owed to anyone for that.

Unlisted is not private. It keeps a duel off every feed, the pool and the recent-duels lists. It keeps it off nothing else: the resolved page publishes both prompts, both images, both provenance records and all five coins at its own URL, anyone holding that URL can read it at any point, it counts in both records, and it pays the jackpot exactly as a public duel does. A duel whose artifacts were never published is not offered at all — it would leave a win nobody could check, and a record with unverifiable entries in it is not a record.

Above 50 a side

A verdict is five coins from a hosted model sampled at temperature 0.4. It can be re-derived from the stored judgment exactly, and it CANNOT be re-run: nobody, including us, can make the same panel say it again. The figures below are what we have measured about how stable that panel is. They are not a guarantee about your match.

• Side preference: 48.1% (was 55.6%), against a band of 45–55%. How often the panel picked whichever fighter it was shown first, measured on mirrored pairs under the flat two-axis weighting. The judges and their instructions are unchanged under rubric-v4.1; the per-coin schedule's own placement bias is structurally unmeasurable from stored runs and is what the season-boundary paired run confirms.
• Split-half reliability: 0.851–0.885 (was 0.865). Agreement between two halves of the same panel over the same fighters, per weight vector: rubric-v4.1 asks a different question each coin, and this is the range across the five questions the schedule asks, from the weight-sweep-e re-scoring of stored runs.
• Coin tie-break rate: 2.6% (was 0.7%). How often a coin's roll-plus-modifier totals tied and fell to the tie ladder — higher modifier first (2.4% of coins), then the side the judge named (0.1%). A duel cannot draw — five coins is an odd number — and no coin is ever redrawn for a tie.
• 5–0 sweep rate: 33% (was 72–86%). Share of matches ending five coins to nil, re-scored from stored runs under the frozen contest configuration (per-coin weight schedule, dice, advantage wheel). Five flat re-rolls of one comparison swept almost always; five contested questions do not, and 3–2 verdicts rose from under a tenth to roughly a third.
• Agreement with a human ranking: 70.2% (was 60.2%). How often the panel ordered two fighters the way a person did, over the calibration roster, under the flat weighting. The schedule deliberately reorders some close matches — that is what it is for — so treat this as a property of the judges, not of any single coin.
• Repeat-pairing disagreement: not measured under the rubric now in force. NOT MEASURED under the rubric now in force. Self-consistency was measured at 96.2% under the previous rubric and is not restated here, because a figure from a rubric that no longer decides anything would be a number with the wrong label on it. Judge the offer on the figures above and on the sentence below them.

Entering above that stake takes a second, explicit confirmation, and we record that these figures were served before your money settles. That is informed consent rather than a warranty: it is the best answer available when the first dispute lands, and it is why the numbers are on this page rather than in a support reply.

What a duel moves

• No seat, no coronation, no defense count.
• No heir. An heir is struck in exactly one circumstance — a throne match resolving SEAT TAKEN over a reigning champion — so a duel breeds nothing. It keeps everything else: stake against stake, a purse, and now sequences.
• No vesting clock and no 47:59 predicate.
• No Elo, no Forge Rating, no Hall of Fame, no fallen-challenger roster.
• No rating of any kind — a Sparring Partner cannot be faked, and a duel opponent can be a wallet you control. A duel record is W–L against DISTINCT counterparties, with the distinct count beside it.

If it goes wrong

Every ending that is not a judged duel refunds in full, and each one carries its own reason code on the public treasury figures. Cancel a listing before anyone takes it and you get everything back — there is no fee, because we incurred no cost and there was no counterparty to harm. A listing nobody takes for 30 days refunds the same way, and never retires your fighter: it means the price was wrong, not that you are finished with the character. If the panel cannot be produced, both stakes go back and no winner is recorded — a walkover is not a win.

Two endings belong to challenges. An invitation nobody answers refunds the inviter under its own code, separate from the pool’s, so a spike in one is legible. And if an arena retires while a duel is waiting for its hour — which a week-long wait makes an ordinary event rather than a theoretical one — both stakes go back in full and nothing is judged.

Not showing up is not a forfeit.If neither side submits actions before the window closes, the arena fills both sequences from a recorded draw and the duel is fought anyway. There is no walkover, no forfeit and no way to win by silence — and the draw is stored, so a disputed loss is never answered with “the server rolled” and nothing to check.

A duel or a challenge?

• A duel is an even match against one other maker, resolved in minutes. The purse is both stakes minus the rake, and it is settled and gone.
• A challenge is a shot at a jackpot that compounds every time anyone else tries, plus the seat, plus forty-eight hours of people coming for it.
• The comparison holds on RETURN, not on size. A $1,000 duel pays $1,800 — you risked $1,000 to net $800. The jackpot is the only prize in the product that can exceed its entry fee by orders of magnitude.