Pearl's Proof-of-Useful-Work makes every "hash" a noised matrix multiplication — the same math that runs every LLM. That is why one GPU fleet can mine PRL, serve inference, and back a yield market.
Any rank-128-ready PRL miner connects over standard pearl/v1 stratum — two minutes, one command. H100/H200 operators can additionally run the official vLLM dual-use miner, which mines PRL as a by-product of serving LLM inference; upstream it is solo-only, and our share-gateway connector for it is in beta. See Connect.
The proof-of-work is a noised matrix multiply: C = (A + noise)(B + noise)ᵗ, with the noise derived deterministically from the block template. Finding a share means finding a computation whose transcript hash beats the share target — there is no separate "hash function" bolted on; the matrix math is the work.
Because the mined GEMMs are real matrix multiplications, inference work can ride the same fleet. That is the basis of our useful-work router and the coming inference-token market: customers submit requests, the router splits them into GEMM tiles, miners compute them, customers get results — and contributors earn a revenue share on top of mined PRL.
Shares are recorded in a crash-proof, idempotent ledger — no double-pays, ever. Payouts go out on the standard PPLNS schedule to your PRL address once you pass the 1 PRL minimum.
How an inference request and a mining share flow through the same system.
Two network upgrades changed what a valid Pearl proof looks like. If your miner predates them, it will look healthy locally while every share is rejected.
Since block 91,630 the network accepts only dense-model certificates; MoE (mixture-of-experts) certificates are rejected. Miners and models built for the earlier MoE fork now produce invalid work.
Since block 96,251 the noise rank must be at least 128, and ranks above 128 are penalized proportionally: the share/block bound is scaled by 128/rank, so rank 256 is twice as hard for the same work. In practice, rank 128 is the operating point. Miners that used higher ranks lost 50-90% of their effective hashrate overnight, which is why every pre-upgrade benchmark table is obsolete.
Consequence: use a rank-128-ready miner version (see the supported miner list), and treat any per-GPU hashrate figure published before August 2026 as unverifiable.