# How Pearl works

https://www.learnpearl.org/how-it-works/ · Checked October 20, 2026 · Beginner

Pearl keeps a Bitcoin-style ledger and changes one big thing: how miners prove their work. Here is the life of a transaction, and why matrix multiplication can stand in for hashing.

## Key takeaways

- Transactions work much like Bitcoin’s: a wallet spends UTXOs, nodes check and relay them, and a miner bundles them into a block.
- Miners multiply matrices with random noise mixed in. The work becomes lottery tickets, and a winning ticket earns the block.
- Each block carries a certificate, a compact zero-knowledge proof, so nodes can check the work without redoing it.
- Difficulty adjusts every block, and the reward shrinks smoothly toward the 2.1 billion PRL cap.

## From transaction to confirmation

_[Diagram: wallet → nodes → miner → matrix work with noise → certificate → confirmed block.]_

1. **A wallet creates a transaction.** It picks spendable outputs, names the new outputs (the payment, plus change back to you) and signs with your key.
2. **Nodes check and relay it.** Every node checks the transaction against the rules before passing it on. Valid ones wait in the mempool.
3. **A miner builds a candidate block.** It picks pending transactions, usually the best-paying ones, and commits to the block’s contents.
4. **Matrix work becomes a certificate.** The miner’s matrix multiplication produces lottery tickets. A winning ticket is turned into a compact block certificate.
5. **The network verifies the block.** Nodes check the transactions, the certificate, the difficulty target and the link to the previous block.
6. **More blocks add confidence.** The block that includes the transaction is its first confirmation, and each later block makes it harder to reverse.

## Why matrix multiplication?

Matrix multiplication is the core arithmetic of AI training and inference: every layer of a model multiplies a grid of inputs by a grid of learned weights. GPUs are built to do it at enormous scale.

Pearl’s thesis is that a network doesn’t need a separate fleet of machines computing throwaway hashes if valuable matrix multiplication can supply the proof of work instead. Try one below: pick any cell of the result to see where it comes from.

_[Interactive: a 3×3 matrix multiplication. Try the full version at https://www.learnpearl.org/tools/matmul-playground/.]_

## How the work is made hard to fake

If a miner could get the answer without doing the multiplication, the proof would be worthless. Pearl’s INT whitepaper, building on a [2025 paper](https://arxiv.org/abs/2504.09971), combines five ingredients to prevent that:

1. **Noise first.** Before multiplying, the miner adds random low-rank noise to both matrices (low-rank means built from a few simple patterns). The noise is derived from the miner’s commitment to its matrices and from the chain, so the miner can’t choose it, and skipping the real multiplication saves nothing.
2. **Commit to the work.** The miner commits to a record of the computation as it happens, so the final answer alone can’t stand in for the work.
3. **Lottery tickets.** Hashes of that committed computation act as lottery tickets. A ticket below the difficulty target wins the block.
4. **Remove the noise.** Because the noise is low-rank, the miner can subtract it cheaply afterwards, and the useful result comes back intact.
5. **Prove it compactly.** The winning work is compressed into a zero-knowledge proof of at most 65,000 bytes that travels with the block as its certificate.

> **Proof of work is not proof of demand.** A valid certificate establishes that the required computation happened. It doesn’t, by itself, establish that a paying AI customer needed that exact computation.

## Consensus and difficulty

Nodes follow the valid chain with the most accumulated work. Pearl aims for one block every 194 seconds. Its WTEMA difficulty rule nudges the target after every block, depending on how long that block took, smoothed over a one-week time constant.

Miners receive newly issued PRL plus transaction fees. When more work joins, blocks come faster and difficulty rises to slow them back down, so a miner’s share of the rewards depends on its work compared with everyone else’s.

- **Target interval:** 194 seconds · Exact in the chain parameters
- **Difficulty:** Every block · WTEMA, one-week time constant
- **Coinbase maturity:** 100 blocks · Before new PRL can be spent
- **Fee market:** First price · Highest fee rate first, no burn

## A smooth emission curve

Pearl caps supply at 2.1 billion PRL. Instead of Bitcoin’s sudden halvings, the reward shrinks a tiny amount with every block. The curve is tuned so about half the supply is issued in the first four years. After that, issuance keeps slowing: the total creeps toward 2.1 billion but never quite reaches it.

**Issued after h blocks:** `issued(h) = 2,100,000,000 × h ÷ (h + 650,226)`

650,226 blocks is four years at 194 seconds per block. Each block’s reward is the difference between two neighboring points on this curve, rounded down to the grain.

_[Interactive chart: Pearl’s issuance curve next to a Bitcoin-style halving schedule. See https://www.learnpearl.org/tools/emission-schedule/.]_

- **Total supply:** 2,100,000,000 PRL
- **Four-year scale:** 650,226 blocks
- **Reward at block 131,829:** 2,232.59 PRL · Chain height on Oct 22, 2026

## How Pearl upgrades

Pearl’s rule changes have so far activated at fixed block heights written into the software, not by miners signaling support. Below a given height one set of rules applies; from that height, the new set does. Operators have to upgrade before the height arrives.

_[Timeline: genesis (block 0), MoE hard fork (71,935), dense-only soft fork (91,630), rank-penalty soft fork (96,251), salted noise-seed hard fork (99,000). See https://www.learnpearl.org/timeline/.]_

## Questions people ask

### Does every Pearl miner have to serve an AI model?

No. The protocol makes mining compatible with real AI workloads, but it can’t force them. Miners can also multiply matrices purely to mine. Useful-work compatibility and actual usefulness are different claims.

### Does Pearl use zero-knowledge proofs?

Yes. Each block’s certificate is a zero-knowledge proof of the matrix work, built with the Plonky2 proving system in the repository’s zk-pow component. Full nodes and the light client both verify it.

### Why 194 seconds?

Pearl’s whitepaper argues that Bitcoin’s ten-minute interval was a cautious 2009 choice, and that modern networks can confirm faster safely. Neither the paper nor the code explains why exactly 194 seconds rather than, say, 180.

### What is FP8, and is it live?

FP8 is the 8-bit floating-point format modern GPUs use for AI. Pearl has published a specification for an FP8 version of its proof, and code for it is being developed on a separate branch. It is not live: mainnet runs the integer scheme with V3 certificates. [Read the FP8 story](https://www.learnpearl.org/news/pearl-fp8-protocol-spec-v4-code/).

## Sources

- [Pearl INT whitepaper (the scheme mainnet runs)](https://pearlresearch.ai/research/int-whitepaper) (Pearl Research Labs, primary)
- [Proofs of Useful Work from Arbitrary Matrix Multiplication (arXiv 2504.09971)](https://arxiv.org/abs/2504.09971) (arXiv, research)
- [Pearl monorepo on GitHub](https://github.com/pearl-research-labs/pearl) (Pearl Research Labs, primary)
- [Proof of Useful Work from the Ground Up](https://pearlresearch.ai/research/blog/proof-of-useful-work) (Pearl Research Labs, primary)