Laplace is the optimistic oracle for AI inference on Solana.
Laplace puts every model answer on Solana with a bond, and anyone can run it again.
CA : 7nB6UdqaY9HihqupRc88nLjpUFos2ivATbe96pjYLAP
first receipt on mainnet is final
asked Market about "solana network fees fall as daily transactions reach a record high". it said negative, 10 s later, for 0.0001 SOL. our watcher ran it again and left a check, nobody challenged, window closed
press Run it again and your tab gets the same bits https://t.co/VGiU38cKCD
how we scale it, roughly in this order
bonds that grow with the money on an answer, through pools anyone can deposit into. and models in the billions of params. both run on our local network now and go live with the next upgrade
after that, agent traces: a whole agent run on one receipt, every call one you can run again https://t.co/bwszc8icP8
the game works at 1.7B parameters now
on our local network a GPU operator answered a SmolLM2 prompt in 197.7M integer steps. a plain CPU watcher matched every bit in about 6 minutes, and a faulty GPU operator lost its dispute in one step on chain
same roots on native, WASM, the chain and CUDA. goes live with the next upgrade
the second model on the network is Market
give it a headline, it says negative, neutral or positive. 33M params, 97.9% on Financial PhraseBank, and every label it posts has 1 SOL behind it for 5 minutes
same headline anywhere, same label, down to the bits
https://t.co/fhQVp2umrj
here's one. a market question goes to Judge, a faulty operator answers it wrong, a watcher's run differs, the game finds the step and the wrong bond goes
a minute, every screen is the real app
markets that already live somewhere else don't have to move. we built a resolver any venue can point at
it settles only on a final judge receipt with enough checks on it. ran it against PNP Markets' program on our local network: the checked yes settled, every wrong path got refused
it goes to mainnet with the first market
agents are paying per call for model answers they have no way to check
an answer someone can run again is the only kind I'd let an agent spend on. so the hold waits for a receipt a watcher actually ran
"nobody's going to bother checking"
a wrong Judge answer has 2 SOL behind it and whoever knocks it down takes half. checking is a browser tab you leave open
https://t.co/taLjtDk1BW
a disputed market on an optimistic oracle goes to a token vote. 48 to 96 hours
fine when the question is about the world. when it's what a model outputs there's nothing to vote on, you run the step
the game we played on our local network took 14 seconds
working on delegated bonds
you deposit behind an operator you think runs the model straight. the pool shares its fees, and if it loses a game the pool eats it too. on our local network one lost game took the share price from 1 to 0.5
it's how a bond gets as big as the money sitting on the answer
an agent's payment can wait on a receipt now
it pays over x402 like it already does. the facilitator puts the money in a hold, and it only moves to the seller once a final, checked receipt shows the answer it paid for. any other answer and it goes back
program's on mainnet: https://t.co/4zLUVEVPbf
"models are random"
the only random-looking part is the sampler, and it draws from a seed. on Laplace that seed is written on the receipt. the rest is integer math
hand someone the dice and they roll your numbers
for a dev it's two lines
ask gets you a receipt with a bond behind it. final resolves once the window closes and nobody knocked it down
https://t.co/g3XkCb3nC3
next on the road: markets that resolve on a receipt
a market names its judge model when it opens. when the judge's answer goes final, it pays out on that. nobody votes
built and tested on our local network, including a resolution that got challenged. the game found the wrong step at 133,316 and the market paid on the right answer. first market on mainnet is next
Market is 1,318,668 integer steps. a browser tab runs all of them in about 11s, Judge in about 1
same bits the operator posted, or it's a challenge
time your own: https://t.co/i94dOALKyU
you can watch the whole network from a browser tab now
pick the models and leave it open. it re-runs every answer, signs a check for each match, and if one differs it challenges and plays the game itself
on our local network it took a wrong Judge answer down in 14s
https://t.co/GTMLUVE0yd
first thing on the road is out: checks
a watcher that re-ran a receipt now proves it on chain. its own key picks a step, it opens the trace there and the program verifies the path. can't fake that without running the whole model
so a receipt shows who actually ran it
https://t.co/CttRVxWR6m
first buyback ran. 2,518,211 LAPLACE burned
20% of the creator fees, 0.2971 SOL, swapped and burned. it runs every hour from here
swap https://t.co/KPQIlDAM9a
burn https://t.co/O6qqcjMbkW