AI Verification Tech MapSearch

Implementation · Cryptographic & computational

Suggest an edit

Pearl proof-of-useful-work blockchain

R3In productionProvider-reported

Pearl is a blockchain developed by Pearl Research Labs whose mining is designed to be a by-product of the matrix multiplications that AI models perform on GPUs.

Each block carries evidence that a certain amount of matrix-multiplication work was done. Its network is live and its code is public. Pearl reports that a mining variant of a 70-billion-parameter model served more tokens per second on four H200 GPUs than the original model, in a benchmark that does not isolate the cost of mining.

An independent study found that the network's dominant mining software appears to multiply random matrices, which verification does not distinguish from AI work. Its R3 (in production) rating is for proving that matrix-multiplication work was done on a blockchain, not for verifying how AI hardware is used. Its security rests on new hardness assumptions, and Pearl lists known shortcuts that let miners do less work than nominal.

Readinesslow confidence

R3 for the narrow claim that GPUs performed a given amount of matrix-multiplication work. The network is live, but no independent security evaluation exists, and the capacity-bounding use in Proofs of useful work for capacity accounting is not demonstrated.

Rubric assessment

Assessed use: proving GPUs did matrix-multiplication work in blockchain mining, not bounding spare capacity

  • R1 met: the protocol, verifier and hardness assumption are specified 1, building on a published construction 5.
  • R2 met: code for a full node, a GPU miner and a proof-of-work circuit and verifier is public 4. Pearl reports running it on H200 GPUs alongside LLM serving 2, against miners who try to win more often than honest work allows 1.
  • R3 met for the narrow claim, as production software that is available and relied on by other parties. Pearl reports that the chain went live when the node code became public 2 4. The repository supports mainnet, and no release up to v1.2.1 is labelled alpha, beta or preview 4. An independent study counted 8,012 online workers among the top 15 miners of one mining pool in May 2026, ran experiments on mainnet, and had shares accepted by that pool, whose acceptance rests on the protocol's verification 6.
  • R4 not met: the README mentions no audit 4, and that study measures how the network is used rather than auditing or attacking the protocol 6. As of September 2026 no independent security evaluation has been published.
Gaps to the next level
  • An independent public security evaluation of the protocol or its implementation.
  • For verification use: a demonstration that proof rates can bound the spare capacity of declared hardware.

Assessed 2026-09-25 against rubric v1.1.

On this page

What it is

Pearl is a layer-1 blockchain developed by Pearl Research Labs. It replaces hash-based mining with a proof of useful work built on matrix multiplication 2 4. The chain launched with an integer scheme 2.

Pearl's September 2026 specification describes a protocol "in which the unit of mining work is FP8 matrix multiplication on GPUs" 1. This extends the integer proof of useful work of Komargodski and Weinstein 1 5. Komargodski is at Pearl Research Labs 7.

Pearl presents the scheme as "2-for-1": GPUs produce proof of work "as a side-effect of AI training and inference workloads" 2. Its public monorepo contains several components 4:

  • a reference full node;
  • a vLLM-based GPU miner;
  • a Plonky2/STARKy-based proof-of-work circuit and verifier.

How it works

Mining. A miner multiplies matrices of its choosing, which Pearl describes as "the operands of an AI training or inference workload" 1. Each block involves four steps 1:

  1. Low-rank noise matrices are derived from public seeds.
  2. The noised product is computed with quantised FP8 arithmetic.
  3. Output tiles are hashed as lottery tickets.
  4. A tile that meets the target wins the block.

The miner then subtracts lower-order correction terms to recover the useful product "at FP8-level accuracy" 1.

Checks on winning tiles. A jackpot policy approves a winning tile only when several statistical checks pass. These checks are meant to exclude degenerate inputs 1.

Verification. The verifier "checks one tile of the matmul, never the whole product" 1. It recomputes that tile bit for bit, reproducing the GPU's rounding and accumulation order off the GPU with the technique of Badash et al. 1 7 (see Deterministic and bit-exact inference). A zero-knowledge verifier can keep the miner's operand data private while proving that the tile satisfies the protocol rules 1. The integer whitepaper describes hash-based SNARKs (Plonky2), compressed to "below 60KB" by recursion 2. An appendix extends verification to mixture-of-experts layers 1.

Evidence

  • A live network. Pearl's integer whitepaper states that the chain became live when the node code was made public 2. The repository supports mainnet and test networks, and has published release v1.2.1 4.
  • A serving benchmark. Pearl reports a benchmark on four H200 GPUs comparing Llama 3.3 70B with its "two-for-one" Pearl-certified variant, which re-implements a layer with a new quantisation mechanism 2:
    • the variant reached 17,206.26 tokens per second with pipeline parallelism, and 18,291.66 with four-way data parallelism;
    • the original bf16 model reached at most 15,269.81 tokens per second, and ran out of memory with four-way data parallelism;
    • MMLU scores were close: 0.8180 to 0.8198 for the variant against 0.8193 to 0.8198 2.
  • Kernel overhead. In a separate benchmark against stock serving engines, Pearl reports 5.08% end-to-end overhead for Llama 70B with four-way data parallelism on four H200 GPUs, and 3.9% for DeepSeek V3.2 with eight-way data parallelism plus expert parallelism on eight H200 GPUs 3. These are developer-reported measurements of its integer proof-of-useful-work kernel.
  • The underlying theory. The construction behind Pearl proves a multiplicative overhead of 1 + o(1) over naive matrix multiplication 5. Pearl's FP8 specification states a "1 + o(1) factor" overhead, without a measured percentage 1.
  • Independent measurement. Basu measured Pearl's mainnet, running the integer scheme, in May 2026, counting 8,012 online workers among the top 15 miners on the AlphaPool mining pool, which Basu estimates to hold about 21% of network hashrate. String analysis suggests that the dominant mining software, from a third party, contains no inference code and generates matrices from random seeds. Random matrices passed verification in the study's tests, and a mining pool accepted shares mined with them 6.

Limitations

Known speedups. Pearl documents known mining speedups and the checks against them 1. It treats faster honest kernels or hardware as outside the threat model 1.

Hardness assumption. Security rests on an informal hardness assumption specific to quantised, noised matrices 1.

Usefulness is not verified. Verification checks the multiplication, not where the matrices came from 6.

Benchmark caveats. The benchmark compares a Pearl-certified model with the original model, and reports no run of the same certified model without mining 2.

Fit to verification. Pearl's proofs show that work was performed, not that no other work ran. Attestable, which proposes proof-of-work accounting for AI agreements, notes that "a proof of some computation is not a proof of all computation", and that bounding spare capacity needs a credible estimate of the compute available 8.

Known flaws

Published flaws, with their severity, kind and status. How flaws are rated.

  • Known mining speedups reduce work per proofMinorTheoretical argumentOpen

    Pearl lists known speedups: crafted inputs, precision shortcuts, seed or commitment grinding, work reuse, and faster kernels or hardware. Its jackpot policy checks limit crafted inputs, and a policy check caps skippable summands at one-sixteenth of those in a tile 1. Pearl describes faster honest kernels or hardware as "not an attack on the protocol" 1.

  • Security rests on a new, informal hardness assumptionSignificantOpen questionOpen

    The FP8 scheme relies on "Assumption 1 (Informal quantized-subspace hardness)": quantised products of noised matrices are assumed not to be substantially easier than generic ones 1. The integer construction it extends lists PoUW from more standard assumptions as an open problem 5.

  • Verification does not check that mined matrices come from AI workloadsMinorTheoretical argumentOpen

    Miners choose their own matrices 5. Basu reports that Pearl's verification "does not check whether the matrices originate from an AI model", that random matrices pass it, and that Pearl's reference mining code generates uniformly random matrices, with vLLM inference as an option. String analysis suggests that the dominant third-party mining software contains no inference code. Basu also finds that a naive fixed-threshold check of matrix kurtosis is defeated, at negligible cost, by sampling clipped Gaussian matrices. Basu calls the gap "a design property" rather than a vulnerability 6. It does not affect the claim that work was performed, but it means the "useful" part of the work is not verified.

Blockers

  • Built for consensus rather than capacity bounding; verifying that declared hardware has no spare capacity would also need a credible compute estimate.

    Capacity bounds[8]

  • Performance figures are provider-reported, and the benchmark reports no baseline of the certified model without mining.

    Adversarial validation[2]

  • Bit-exact verification depends on reproducing GPU arithmetic deterministically.

    Performance & compatibility[1]

Sources

  1. BPearl Research Team (2026). Pearl Floating Point Scheme Specification. Pearl Research Labs. Source recordSupports: FP8 protocol, verifier, jackpot policy, known speedups, ZK verifier, assumption (provider-reported) · abstract; §1; §2; §4.1; §5.1; §6; §7; App. A–B
  2. BPearl Research Labs (2026). Pearl INT Whitepaper. Pearl Research Labs. Source recordSupports: INT scheme, launch statement, benchmark table, zkSNARK size (provider-reported) · §3; §4.2; §4.7; §7 Table 1
  3. CPearl Research Labs (2026). Proof of Useful Work from the Ground Up. Pearl Research Labs. Source recordSupports: provider-reported end-to-end serving overhead for the integer proof-of-useful-work kernel · Measured in production
  4. BPearl Research Labs (2026). pearl: Monorepo for the Pearl network. GitHub. Source recordSupports: public node, miner and ZK prover code; releases · README; release v1.2.1 (commit e72ad36)
  5. BI. Komargodski & O. Weinstein (2025). Proofs of Useful Work from Arbitrary Matrix Multiplication. arXiv. Source recordSupports: underlying PoUW construction and its assumptions; miners choose the instance · abstract; §2; §3
  6. BA. Basu (2026). The Usefulness Gap in Proof-of-Useful-Work: An Empirical Study of Pearl's cuPOW Protocol. arXiv. Source recordSupports: independent measurement of Pearl's mainnet (integer scheme): worker count, mining software, random-matrix mining, verification scope, statistical check defeated · abstract; measurement and verification sections; conclusion (v2)
  7. AE. Badash et al. (2026). Hawkeye: Reproducing GPU-Level Non-Determinism. Proceedings of Machine Learning and Systems 8 (MLSys 2026). Source recordSupports: Hawkeye, the bit-exact GPU reproduction the FP8 verifier uses; Komargodski's affiliation with Pearl Research Labs · author block; abstract
  8. CAttestable (2026). Pacing AI Requires Proof. Attestable blog. Source recordSupports: a proof of some computation is not a proof of all computation; capacity bounding needs a compute estimate (Attestable's proposal)

Search

Full search page