zorch

Composable building blocks for SNARKs, running on FRX — Fractalyze's fork of JAX with native finite-field dtypes. You assemble a prover the way deep learning stacks layers: reusable stages, chained until nothing is left to prove.

Install

Python 3.11, Linux x86_64 or macOS Apple Silicon. Install as pyzorch, import as zorch:

pip install pyzorch                    # CPU — nothing else needed
pip install pyzorch 'frx[cuda12]' \
    --extra-index-url https://fractalyze.github.io/pypi/simple/   # GPU (CUDA 12)
python -c "import frx, zorch; print(frx.devices()); print(zorch.__version__)"

Platform limits and error symptoms: Setup.

The mental model

A proof system is a chain of claim reductions. Each stage has paired prover/verifier roles; both derive the same reduced claim, and the last claim holds by construction:

flowchart LR
    C0["SumClaim<br/><i>“f·h sums to v”</i>"] -->|"SumcheckProver<br/>(one round per variable)"| C1["EvaluationClaim<br/><i>“…if f·h(r) = u”</i>"]
    C1 -->|"PCS opening stage"| C2["TrivialClaim<br/><i>holds by construction</i>"]

Inside a stage, rounds repeat one recurrence — a message observed into the Fiat-Shamir transcript, a challenge sampled back — and the transcript is a device-side value threaded through every call:

sequenceDiagram
    participant P as ProverRound
    participant T as Transcript (device)
    participant V as VerifierRound
    P->>T: observe(message)
    T->>P: sample(challenge)
    Note over P: fold state with challenge
    V->>T: observe(same message)
    T->>V: sample(same challenge)
    Note over V: replay, check, carry RunningClaim

Find your door

• Get started — learn by doing

  ---

  Install, prove a sumcheck, verify it, and read what happened — about two minutes, CPU only.

  First proof

• Guides — get a task done

  ---

  Assemble a full prover, keep your code fused on GPU, fix install and toolchain issues.

  Assembling a prover

• Reference — look it up

  ---

  Every block's import path with worked examples, the field-dtype sharp bits, and the generated API.

  Blocks & imports

• Design — understand the WHY

  ---

  The load-bearing decisions on one page — the two non-negotiables, the fusion contract, one line per block. Full prose lives on GitHub.

  Design digest

Using an AI coding agent?

The Guide section doubles as an installable agent skill: npx skills add fractalyze/zorch gives your agent the same pages.