Proof of Useful Work (PoUW) is Pearl's consensus mechanism. Miners prove they did real computational work — as every proof-of-work chain requires — but the work they do is matrix multiplication, the operation at the heart of AI inference and training.
The result is the same computation serving two purposes: producing a useful AI result, and securing the blockchain.
Why this was thought impossible
Proof of work needs a puzzle with specific properties. It has to be hard in a predictable way, easy to verify, and impossible to shortcut. Bitcoin achieves this with hashing — deliberately arbitrary work whose only value is being hard.
Useful computation breaks those properties. If the work is something a miner wants to do, they can choose easy inputs. If it produces a valuable result, they can reuse it. For years the consensus was that useful work couldn't be secure work.
How Pearl gets around it
The trick is adding noise, then removing it.
- Start with a real workload. Two matrices, A and B, from genuine AI computation. The useful result is A·B.
- Add cryptographic noise. Derived from the current blockchain state, the protocol generates noise matrices and computes (A+E)·(B+F) instead. Because the noise depends on chain state the miner can't control, the noised product is as hard to compute as multiplying two random matrices — so choosing easy inputs doesn't help.
- Extract a proof from the work. As the multiplication runs, the protocol hashes intermediate values from the actual execution. Occasionally one of those hashes falls below the difficulty target — that's a winning block, exactly as a Bitcoin miner occasionally finds a qualifying hash.
- Recover the useful answer. The noise is low-rank by construction, which means A·B can be recovered from (A+E)·(B+F) with a cheap correction step. The miner gets their AI result; the network gets its proof of work.
Why miners can't cheat
Fake inputs don't help. The noise makes the computation hard regardless of what A and B are. Multiplying matrices of zeros is no faster.
Shortcuts don't work. A miner must submit not just the result but a verifiable trace of intermediate values from the real computation. This proves the multiplication was performed directly rather than reconstructed from a precomputed answer.
Verification is cheap. Checking a proof requires far less work than producing it — the standard asymmetry proof of work depends on. Multiplying t×t matrices takes on the order of t³ operations while the hashing overhead is only t², so the added cost of mining is negligible relative to the computation itself.
Your workload stays private
Miners commit to their matrices cryptographically rather than publishing them, so the underlying data is never revealed. Pearl layers a zero-knowledge proof on top, so a miner can prove they did valid work without disclosing anything about the computation. The proofs are hash-based, which means they're expected to remain secure even against quantum adversaries, and they need no trusted setup.
Does mining slow down my AI work?
The mining kernel is designed to interleave with normal AI computation at negligible overhead. In Pearl's own published benchmarks, the Pearl-certified variant of Llama 3.3 70B matched the original model's accuracy while producing mining work — in some parallelism configurations it ran faster than the baseline.
See What is the Pearl vLLM miner? for how this works in practice.
Is it ASIC-resistant?
Pearl's argument here is economic rather than cryptographic. An ASIC purpose-built to mine Pearl would lose the ability to do general useful work — and since Pearl mining draws its value from doing useful computation, that specialised chip ends up less profitable than a general-purpose GPU doing both jobs at once.
What's next for the protocol
The current scheme performs exact integer matrix multiplication. A planned upgrade moves to high-accuracy approximate multiplication, using the quantisation noise already present in low-precision AI workloads as the perturbation mechanism. That would let Pearl support training workloads and state-of-the-art low-precision inference without adaptation.
Read more
The full mechanism, with algorithms and security analysis, is in the Pearl whitepaper, Sections 3 and 4. The academic paper is at arXiv:2504.09971.