Snapverdict markSnapverdict
Scored on chain. The vault factory is on BNB Chain, and a full season ran end to end on BNB Chain testnet: sixteen answers, every one scored by the contract.
A pack of 128 judgement items drifting toward a deadline line; past it, a single answer card being stamped with one action.one pack128 items · one policy carddeadlineone action per item
A pack is 128 short judgement items sharing one policy card. One request answers all of them; the block the answer lands in decides whether it counts.

A pack of 128 judgements, answered within two blocks.

Snapverdict is a meme token on BNB Chain with one job: take its share of the 2% tax on its own trading (the part Flap forwards), hold it in a vault as BNB, and hand it to whoever answers the round's questions correctly and on time. There is no treasury behind it and nothing is minted. Without trading, the pool only shrinks.

128

items in a pack

one shared policy card, one action each

2

blocks to answer

from the opening, and within 4 s of the beacon; measured p95 is 344 ms

2%

of every trade

the only source of rewards, paid in BNB

6h

per round

credit × your stake, scored by the contract

why a fast model, and not a smart one

Two gates, and both of them are measured.

The questions are deliberately the shape one request answers in parallel: short, typed, independent, sharing a single policy card. That shape is what makes the first gate possible. A chat model has to fan the pack out across branches, pay for the policy card on every branch, and wait for its slowest one.

Four systems running at two gates in series: a clock gate of two blocks within four seconds of the beacon, and a whole-pack budget of two errors in 128. Contender B stops at the clock; Contenders A and C stop at the error budget; only Jev's lane reaches the end.two gates, in seriesthe clock2 blocks · 4 sthe whole pack≤ 2 wrong in 128JevContender BContender CContender Aa pack that misses either gate scores zero, not less
The clock keeps out the slower timed contender; the faster one is within tens of milliseconds of the window, and the whole-pack error budget stops most of its packs.

The clock

Measured from a US machine on a pack of 128: p50 318 ms, p95 344 ms, and 100% inside 400 ms. The nearest contender needed 1061 ms and the most accurate one 1958 ms — 0% on time, both of them.

The hard items

On the 90 hardest items in the sample, the three contenders were wrong on 29.2% to 77.8% of items. Jev was wrong on 0. On the whole screened pool the contenders were wrong on 3.6% to 7.4% of items; in the shipped season (128 items, two errors allowed) that is about 15% of packs passing for the best contender and under 2% for the others.

Honest about the shape of that second number: the entire accuracy gate rests on 90 items. It is the thinnest part of this project, and it is written up as such on what was measured.

the loop

What happens every ten minutes.

  1. 01The round is committedBefore a round starts, the vault is told each pack's ciphertext hash, a commitment to its answer key, and the public randomness beacon round that opens it. None of it can change afterwards.
  2. 02Every rig has the pack earlyThe sealed pack is published minutes ahead; each rig downloads it and checks it against the hash on chain. At the beacon instant nothing is sent, so nobody wins on bandwidth.
  3. 03The beacon firesThe rig opens the pack with the beacon, puts the signature on chain (the contract verifies it itself), asks the model, and commits its answers within two blocks of the opening and four seconds of the beacon.
  4. 04The contract scores itThe key is opened after the window. The contract scores every answer against it and weighs the score by the wallet's stake. After the season, you claim your own BNB.

The instant a pack opens is picked by the operator and committed on chain before the round, at least a minute ahead; after that nobody can move it. What opens the pack is a public randomness beacon's signature for that instant, which the contract checks itself and which nobody, including whoever runs this, can know before the beacon produces it. At that instant not a single byte is sent, so nobody wins on bandwidth.

state of play

Where things stand.

Vault factory
On BNB Chain at 0xEde8210927618393F3FA5ff3F0F5A7ae8A0c95fF. Its vaults score every answer on chain; the upgrade key of the beacon behind them belongs to Flap's Guardian. Only an address published on this page is the token's; treat any other as not ours.
Testnet season
Run end to end on BNB Chain testnet on 25 September: a token launched through the factory, one real buy paying the tax in, two seasons, eight packs, two rigs in Tokyo answering with Jev. All sixteen answers landed inside the window and the contract scored every one of them in full; the rigs then claimed their BNB, split by stake, and took their stakes back.
Rig
One Node.js file. It stakes, self-tests against the window, and refuses to mine on a machine that cannot make it.
What is measured
The two gates: how fast a pack comes back, and who gets it right. Those numbers are real and reproducible.

six pages

Read the rest.

How it works One card, 128 tickets, one action each — and the clock that starts on a public beacon.Run a rig Five steps. A stake, one small instance near the model's API, your own keys, and a self-test.Rules and parameters Every number you need to recompute your own score, with the reason it is that number.What was measured Two gates, measured on real runs — and a plain list of what was not measured.Rounds and logs What a round publishes, how the contract pays you, and the leaderboard.Questions Affiliation, keys, regions, earnings, and what happens when a label is wrong.