Loading
One coin, launched on two chains in the same second. Every vamp of it joins the same meme. Half an hour later, at a time nobody could have predicted, one of them takes the pot, and keeps taking it.
Every launch request is compared against everything launched on the platform in the last 30 minutes, plus the medoid of every meme with an open window.
Five cheap, deterministic signals run first: the X handle and website canonicalized and matched exactly, the image compared by perceptual hash and by CLIP similarity, the ticker by Jaro-Winkler after de-leeting and stripping $, coin, inu, token, the name by normalized edit distance, and the deployer by wallet or funder within two hops.
Those fuse into a single probability. Confident matches merge and confident non-matches open a new meme without ever calling a model; only the uncertain band goes to a Claude adjudicator, which sees both images and the normalized text and returns a structured verdict. The comparison is against each cluster’s medoid, so the cost stays linear in the number of memes, not quadratic in the number of coins.
A vamp is not blocked. It is routed: it joins the original’s meme id, and its fees join the original’s pool.
One submission deploys on pump.fun (Solana) and PONS (Robinhood Chain) in the same second, at the same opening cap, with one supply per chain.
A per-coin market maker holds seed inventory and cash on both chains. When one side runs more than 5% ahead of the other it sells into that side and buys the cheaper one. It is not target-seeking and it does not defend a price; it only ever leans against the gap, so the two markets stay within arm’s length of each other without anyone underwriting a peg.
On Solana the creator vault is derived from a platform-held key that is unique per meme, so one sweep collects the creator fee from every coin in the meme. On PONS the same job is done by setting the meme’s fee pool as the creatorFeeRecipient. The pooling is an on-chain fact on both sides, not an accounting convention.
Every ten seconds, each coin in the meme is scored against the others in its cohort, on a ranking designed so that faking it costs more than earning it.
The score is a weighted geometric mean of four log-scaled, cohort-normalized inputs: market cap, fees actually paid, wash-adjusted volume and effective holders. Geometric, because it is non-compensatory: inflating one input tenfold at weight 0.2 buys a factor of 1.58, where an average would have paid about five. The cost of gaming grows exponentially as the weight of what you are gaming falls.
Five multiplicative penalties then apply, each in (0, 1]: holder concentration, Benford deviation on trade-size leading digits, trade-size entropy, net flow versus gross (a wash loop drives this to nothing), and same-slot bundled buys at launch. The result is smoothed with a five-minute half-life, nudged by cohort-relative momentum, and shrunk toward the cohort median while the trade count is thin, so a coin with nine trades cannot leapfrog one with nine hundred.
A second, independent ranking, Kemeny-Young over the four single-metric orders, runs alongside it. When the two disagree on the top five, the meme is flagged disputed on its page rather than quietly showing one of them.
The close time is random, committed in advance, hidden until it happens, and verifiable by anyone afterwards.
At meme creation the settler draws a close time uniformly from a four-minute range, generates a salt, publishes sha256(close ‖ salt) and encrypts the time itself under a key only it holds. The countdown you see on a meme page is the advertised close, not the real one.
When the window closes, the time and the salt are revealed. Anyone can hash them and check the result against the commitment that was public half an hour earlier, which is what makes last-block sniping pointless: there is no way to know which block is the last one, and no way for the settler to move the deadline after the fact.
The winner is the coin with the highest score at the last tick at or before the close. That is a deterministic function of data that was already published, so the result can be replayed from the tick history by anyone who kept it.
Every creator fee from every coin in the meme flows into one pool, before the close and forever after it.
At the close the pool accumulated during the window, and every fee unit that arrives after it, is split the same five ways. The 10% maker reserve rolls into the burn once the reserve is full, so in the long run 70% of the pool is spent buying the winner and destroying it.
Buy the meme, not the coin. One order with SOL or ETH spreads across the meme's eligible top five at the weights of the tick it was quoted in.
The weights are a clipped softmax over the live scores, water-filled back to 100% after clipping, which makes them Plackett-Luce win probabilities. An even race lands near 40/25/17/11/7; a near-tie between the top two flattens to about 35/34/15/10/6; a blowout pushes the leader toward the 45% cap and no further. The floor exists so rank five is never dust, and the cap so one coin is never simply “the bucket”.
The rates are the chains’ own where they are fixed, and ours where they are not. Nothing here is hidden in a tooltip on a confirmation screen.
Nothing stops it from launching. What changes is where it lands. A launch inside the 30-minute window that matches an open meme on its X link, website, image, ticker or name joins that meme's id instead of starting a new one. Its creator fees flow into the same pool, and it competes for the same pot.
Cheap deterministic features run first and fuse into one probability: canonicalized links, perceptual image hashes, string distance on ticker and name, deployer and funder overlap. Above 0.90 it merges automatically, below 0.40 it opens a new meme, and only the uncertain band in between is sent to a Claude adjudicator with both images. An operator can override any assignment; every override is appended to the audit log.
Nobody knows, which is the point. The close time is drawn uniformly from 28 to 32 minutes and encrypted at meme creation; only the hash of it is published while the window is open. At close, the time and salt are revealed, and anyone can recompute the hash and check it against the one that was published half an hour earlier. A publicly known deadline would just be a sniping target.
It is deliberately expensive. The score is a weighted geometric mean, not an average, so inflating one input ten-fold at weight 0.2 buys you 1.58x, not 5x. Wash trades are dropped before volume is counted, and multiplicative penalties cut the score for holder concentration, Benford deviation in trade sizes, low trade-size entropy, near-zero net flow and same-slot bundled buys at launch.
One order, spread across the meme's eligible top five at that tick's weights, held in custody. The weights are a clipped softmax over the live scores, which makes them Plackett-Luce win probabilities, so a bucket is, honestly stated, a probability-weighted claim on whichever coin wins. You sell back into the asset you paid with.
They keep trading; nothing is confiscated. What their creators stop receiving is the pool: from the close onward, every fee unit from every token in the meme is split five ways that do not include them. The winning coin is bought and burned with 60% of it, forever.
Once per wallet per 24 hours. A launch request waits up to 48 hours for approval and, once approved, 24 hours for payment before it expires. One submission deploys on both chains in the same second.
One submission, both chains, one meme. If your idea already has an origin, Prism routes you into its meme instead of splitting the attention.
Launch a coin