Pearl research library

1. 01

   The INT whitepaper

   The scheme mainnet runs: protocol design, the proof-of-useful-work construction, parameters, addresses and issuance. Read the INT whitepaper (opens in a new tab).

2. 02

   Proofs of useful work from arbitrary matrix multiplication

   The 2025 paper by Ilan Komargodski and Omri Weinstein that the repository names as Pearl’s basis: the math that lets chosen matrix multiplication serve as proof of work. arXiv 2504.09971 (opens in a new tab).

3. 03

   The open-source implementation

   What actually ships: pearld, Oyster, the light client, the proof system and the vLLM miner. GitHub (opens in a new tab).

4. 04

   The usefulness-gap study

   An empirical challenge to whether observed mining was useful AI work, measured in late May 2026. arXiv 2606.04819 (opens in a new tab).

5. 05

   Pearl’s explainer and benchmarks

   Pearl’s own walk-through of its approach, with its measured serving overhead. Proof of Useful Work from the Ground Up (opens in a new tab).

6. 06

   The economics

   When block rewards can increase the supply of useful compute without making attacks cheap. arXiv 2606.06700 (opens in a new tab).

7. 07

   An outside systematization

   An independent reading of Pearl’s proof-of-useful-work scheme that proposes letting each miner choose its own noise level. arXiv 2610.25347 (opens in a new tab), October 2026.

8. 08

   What comes next: FP8

   “The Pearl Protocol” specifies an FP8 floating-point version (not live), drawing on Hawkeye (opens in a new tab) for reproducing GPU arithmetic exactly. Read the spec (opens in a new tab).

Cryptographic

Can MatMul prove work?

Verification and security

Technical

Can it share an AI workload?

Integration and overhead

Empirical

Is current mining useful?

Observed real-world behavior

Economic

Will useful compute create value?

Adoption, incentives and demand

• What fraction of network work is paired with real training or inference?
• How concentrated are miners, pools, proof generation and AI customers?
• What proof overhead and verification cost show up across current model architectures?
• How does the move to FP8 (certificate V4) change accuracy, hardware requirements and security?
• How does the network behave under fast swings in work and adversarial timestamps?
• Can wallets and exchanges mature without becoming central points of failure?

Protocol facts come from the code and the official whitepapers, and where the two disagree, the code wins. Release facts come from the project’s GitHub releases. Performance and adoption claims are labeled by source and date. Third-party tutorials are never treated as protocol specifications.

When credible sources disagree, we say so and name the evidence that would settle it. A project claim doesn’t become a fact by being repeated.