# Pearl publishes an FP8 protocol spec, and code for certificate V4 lands. It isn’t live.

https://www.learnpearl.org/news/pearl-fp8-protocol-spec-v4-code/ · October 5, 2026 · Research

The work would move onto the number format GPUs actually run AI in. For now, mainnet still runs the integer scheme.

## What happened

Pearl Research Labs’ site now presents a September 2026 document, the “Pearl Floating Point Scheme Specification”, as its main whitepaper under the title “The Pearl Protocol”. The original integer-scheme paper stays online as the [INT whitepaper](https://pearlresearch.ai/research/int-whitepaper), and that is the scheme mainnet runs.

Since September 9, fp8-scheme, a development branch of the repository, has carried code for a fourth certificate version, which the code calls the FP8 scheme. Pull requests merged there between October 2 and 5 added verification work ([#355](https://github.com/pearl-research-labs/pearl/pull/355), [#363](https://github.com/pearl-research-labs/pearl/pull/363)) and split the proof code into separate v1, v2 and v4 modules ([#362](https://github.com/pearl-research-labs/pearl/pull/362)). None of it is on the master branch the releases are built from, and no mainnet fork height exists for it.

## Integer today, FP8 specified

| Feature | INT scheme (live) | FP8 scheme (specified) |
| --- | --- | --- |
| Arithmetic | Exact integer matrix multiplication | FP8 floating point, the format GPUs use for AI |
| Lottery tickets | Hashes of the computation’s transcript | Hashes of output tiles |
| How a verifier checks | Recomputes exactly in integers | Recomputes one tile bit for bit, reproducing the GPU’s own arithmetic |
| Hardware rule | None in consensus | A whitelist of GPU architectures, named in each proof |
| Certificate | V3 since block 99,000 | V4, in development |

## Why it matters

Modern AI inference increasingly runs in low-precision floating point. A proof of work that runs natively in FP8 could fit real workloads more closely than an integer path that needs separate quantization.

The spec is also candid about its limits: its central hardness property is stated as an assumption rather than proven, and its rules cap, rather than eliminate, the speed-up a miner that doesn’t need the matrix product can gain over an honest one.

## How to read it

> **Specified is not shipped.** Say “Pearl has published an FP8 specification and code is landing on GitHub”, not “Pearl now mines in FP8”. There is no launch date, and the code has already moved past details in the published PDF, so expect revisions.

## Sources

- [“The Pearl Protocol”: the FP8 specification (not live)](https://pearlresearch.ai/research/whitepaper) (Pearl Research Labs, primary)
- [Pearl INT whitepaper (the scheme mainnet runs)](https://pearlresearch.ai/research/int-whitepaper) (Pearl Research Labs, primary)
- [PR #355: FP8 (certificate V4) verification work](https://github.com/pearl-research-labs/pearl/pull/355) (Pearl Research Labs, primary)
- [PR #362: zk-pow split into v1, v2 and v4 modules](https://github.com/pearl-research-labs/pearl/pull/362) (Pearl Research Labs, primary)
- [Hawkeye: Reproducing GPU-Level Non-Determinism (arXiv 2603.20421), by researchers at Pearl Research Labs and Stanford](https://arxiv.org/abs/2603.20421) (arXiv, research)