Start

What Laplace is

Receipts, bonds, windows and the people on the network, in one page.

Laplace is a network on Solana where a model's answer is posted as a receipt with a bond behind it, anyone can run the model again, and a disagreement is settled by the program executing one step. This page names each part in a sentence or two and points to the page that covers it.

  1. 1requestfee into escrow
  2. 2answerbond locks, window opens
  3. 3windowanyone runs it again
  4. 4finalizefee paid, bond unlocks
A request from Ask to final.

The receipt

A receipt is one account of the laplace program. It holds a request (the model, the input, the seed, the fee) and, once the operator answers, the output, the trace root and the bond locked behind them. It moves through five states: requested answered challenged final void.

Receipts lists every field and what moves a receipt from one state to the next.

The bond and the window

Every answer has SOL locked behind it for a set time. The amount is the model's bond and the time is the model's window. Both are set by the model's publisher when the model is registered, and they are the same for every answer to that model. During the window anyone can challenge the answer. An answerer that loses a challenge loses the bond.

Bonds and windows covers where a bond sits, what moves it and what pauses a window.

Run it again

A model on Laplace is an integer program, so every machine that runs it on the same input and seed gets the same bits: the same output and the same trace root. Run it again on a receipt's page runs the model in the visitor's browser and shows its bits beside the operator's, as match (≡) or differ (≢). It takes no wallet and sends nothing.

Deterministic inference explains why the bits are the same everywhere. Run it again yourself does it in the app and with the SDK.

The challenge and its one step

A run that gives a different output can be posted against the answer, with a bond equal to the model's. That is a challenge, and it opens a dispute. The two sides then play the game: turn by turn they narrow the disagreement between their two runs to a single step of the model. The program executes that step itself and compares the result with what was claimed. The side that was wrong loses its bond, and the answer that was right stands on the receipt.

The challenge game walks the turns. One step on chain covers what the program executes and what the transaction carries.

Final

When the window closes with no dispute open, anyone can finalize the receipt. The fee leaves escrow to the answerer, less the protocol's share, the bond unlocks, and the state is final. A final receipt is what other software acts on: a Solana program reads it through the laplace-receipt crate, which returns the output only in that state, and a client waits for it with laplace.final(receipt).

Fees and offers covers the fee. Quickstart: read a receipt from a program reads one on chain.

The people on the network

Four kinds of person use Laplace. Each has a page to start from.

RoleDoes
RequesterPicks a model and an operator's offer, sends the input and pays the offer's price. A person does it in the app, a developer or an agent through the SDK, a program by reading the receipt.
OperatorRuns the model and posts the output with a bond. Is paid the fee, less the protocol's share, when the receipt is final.
WatcherRuns the model again and challenges when the bits differ. Wins part of the bond of an answer that was wrong.
PublisherRegisters a model once, with its window and its bond.