jolt

name	jolt
description	Wrap a Rust function in a Jolt zero-knowledge proof
allowed-tools	Bash, Read, Edit, Write, Glob, Grep, Task

Invoke when the user says: "make this Jolt provable", "wrap this in Jolt", "prove this with Jolt", "add ZK proofs to this", "make this zero-knowledge", "make this provable", "jolt-ify this".

Step 1 — Identify the computation to prove

Look for a pure, deterministic Rust function — inputs in, result out, no I/O or side effects. If not obvious, ask:

"What function should I make provable? It needs to be a pure Rust function with no I/O or side effects."

Before writing any guest code, verify the target function and its entire module path are pub. If not, make it pub in the library source (preferred — we're proving the library) and confirm with the user, noting that inlining is an alternative if they'd rather not modify the library.

Step 2 — Analyze and adapt the signature

The guest has a real heap — Vec, String, alloc types work freely inside the body. The constraint is at the parameter boundary: std mode uses full serde (Vec/String as params fine); no_std uses serde_core (no Vec params, arrays capped at size 32). Only adapt what's necessary:

Issue	Resolution
Vec<T> param in no_std	[T; N], len: u32 — or switch to std mode
[T; N] where N > 32 in no_std	Split across multiple params (serde_core array size limit)
usize	u64 (guest is 64-bit RISC-V)
f32 / f64	Fixed-point integer (e.g. i64 * 1_000_000) — RV64IMAC has no FPU
std::io, std::net	Cannot run in guest — explain and stop
Non-determinism	Pass seed/timestamp as explicit input

Build mode: read the library's Cargo.toml. Use std mode if the library requires std, or if it makes the example simpler (e.g. Vec/String as params). No_std is a choice, not the default.

Step 3 — Install Jolt

jolt --version  # check if installed
cargo install --git https://github.com/a16z/jolt --force jolt  # if not

Step 4 — Scaffold

If inside an existing Rust library repo, propose:

"I'll create <library-name>-jolt/ here with the proof scaffold and import your library as a path dependency. Sound good?"

jolt new <project-name>        # standard mode
jolt new <project-name> --zk   # with PrivateInput + BlindFold support

This generates a workspace with a fib example — replace it by renaming fib → <fn> throughout src/main.rs and guest/src/lib.rs. Preserve the [patch.crates-io] block in the root Cargo.toml (required arkworks patches).

Critical: always use jolt new to scaffold. The generated guest/src/main.rs (#![no_main] binary stub that re-exports the lib) is required for Jolt to produce a RISC-V ELF binary. Without it, cargo builds only a .rlib and the host panics with "Built ELF not found." If creating the guest manually, always include:

// guest/src/main.rs
#![cfg_attr(feature = "guest", no_std)]
#![no_main]

#[allow(unused_imports)]
use guest::*;

Step 5 — Write the guest (guest/src/lib.rs)

no_std mode (default):

#![cfg_attr(feature = "guest", no_std)]
extern crate alloc;  // heap always available

#[jolt::provable]
fn <fn>(<params>) -> <ret> { ... }

std mode — in guest/Cargo.toml:

jolt = { package = "jolt-sdk", git = "https://github.com/a16z/jolt", features = ["guest-std", "thread", "stdout"] }

Include "thread" for rayon/parallel, "stdout" for println!. No cfg_attr needed in the lib file.

Macro parameters — use #[jolt::provable] bare; only add parameters when you have a reason:

Parameter	Default	When to change	How to pick a value
stack_size	4096	stack overflow	Start at 8388608 (8 MB, matches Linux default); reduce in the optimization pass.
max_trace_length	2^22	max_trace_length exceeded	Run analyze_<fn> to get actual cycle count, round up to next power of 2. Proving time and memory scale with this — tighten in Step 9.
heap_size	32 MB	heap allocation failed	Estimate peak live allocations; halve until it fails, then double back.

Prover-only inputs — two options depending on whether you need cryptographic privacy:

• jolt::UntrustedAdvice<T> — prover-only; excluded from the verifier API but values may be recoverable from the proof
• jolt::PrivateInput<T> — same underlying type, signals that values should be cryptographically hidden via BlindFold (requires zk on the host, not the guest)

#[jolt::provable]
fn my_fn(public: u64, secret: jolt::UntrustedAdvice<[u8; 32]>) -> bool {
    let secret = *secret;
    // ...
}

Host prove call: prove(..., UntrustedAdvice::new(val)). The generated verifier signature omits the advice entirely. Add use jolt_sdk::UntrustedAdvice; to the host.

For PrivateInput<T>, enable zk on the host only (see Step 7). The macro enforces this at compile time.

TrustedAdvice<T> is the alternative for data committed by a third party — it requires a commit_trusted_advice_<fn>(...) host call and the commitment is passed to the verifier.

Dependencies — add to guest/Cargo.toml. When wrapping an existing repo, add <library> = { path = "../.." }. Avoid default-features = false unless you know the library supports it — disabled default features can expose conditionally-compiled modules that still reference missing optional deps. For crypto, prefer jolt-inlines-sha2, jolt-inlines-keccak256, jolt-inlines-secp256k1.

Multiple functions — each #[jolt::provable] generates independent compile_*, preprocess_*, build_prover_*, build_verifier_* APIs.

Advice functions — for expensive witness computation that should run outside the proof, use #[jolt::advice] in the guest. The function runs on the host/prover; the guest verifies the result cheaply with jolt::check_advice!(bool_expr) or jolt::check_advice_eq!(a, b).

Advice-based modular exponentiation — for RSA/bigint modexp, the host computes (quotient, remainder) for each modular multiply via #[jolt::advice]; the guest only verifies a*b == q*n + r (two schoolbook wide multiplies + equality check) and r < n. This drops RSA-2048 verify from ~12M to ~461K cycles (26x). For e=65537 (2^16+1): 16 squarings + 1 multiply = 17 advice-verified steps. This pattern works for any modular exponentiation.

Hash inline performance — cycle counts at 64-byte input (with inline acceleration):

Hash	Cycles	vs SHA-256	Max input
BLAKE3	863	8.2x faster	64 bytes (single block only)
Blake2b	1,264	5.6x faster	unlimited
Keccak-256	3,680	1.9x faster	unlimited
SHA-256	7,134	baseline	unlimited

BLAKE3 is fastest but capped at 64 bytes. For variable-length hashing, Blake2b is the best choice. SHA-256 must be used when the hash is dictated by an external protocol (e.g., JWT RSA-SHA256 signatures).

Cycle tracking — instrument sections of the guest to measure per-section cycle counts (visible in the prover log):

use jolt::{start_cycle_tracking, end_cycle_tracking};

start_cycle_tracking("my section");
// ... code to measure ...
end_cycle_tracking("my section");

Step 6 — Write the host (src/main.rs)

use std::time::Instant;
use tracing::info;

pub fn main() {
    tracing_subscriber::fmt().with_env_filter(
        tracing_subscriber::EnvFilter::from_default_env()
    ).init();

    let target_dir = "/tmp/jolt-guest-targets";
    let mut program = guest::compile_<fn>(target_dir);
    let shared = guest::preprocess_shared_<fn>(&mut program);
    let prover_prep = guest::preprocess_prover_<fn>(shared.clone());
    let verifier_setup = prover_prep.generators.to_verifier_setup();
    let verifier_prep = guest::preprocess_verifier_<fn>(shared, verifier_setup, None);
    let prove = guest::build_prover_<fn>(program, prover_prep);
    let verify = guest::build_verifier_<fn>(verifier_prep);

    let t = Instant::now();
    let (output, proof, io) = prove(<inputs>);
    info!("Prover runtime: {} s", t.elapsed().as_secs_f64());

    // io.panic is true if the guest panicked; the verifier checks it matches the proof
    let is_valid = verify(<inputs>, output, io.panic, proof);
    info!("output: {:?}", output);
    info!("valid: {is_valid}");
    assert!(is_valid);
}

For multiple functions, replicate the block per function. To measure cycles before proving: guest::analyze_<fn>(<inputs>).write_to_file("summary.txt".into()).unwrap().

Step 7 — Run

Before running, estimate peak memory from max_trace_length (conservative worst-case):

max_trace_length	Peak memory
≤ 2^23	< 10 GB
2^24	~15 GB
2^25	~32 GB
2^26	~42 GB
2^27	~81 GB
2^28	~99 GB

If max_trace_length is 2^24 or above, warn the user and ask how to proceed:

"This may require ~X GB of RAM. I can: (a) run analyze_<fn> first to get the actual cycle count — if it's well below max_trace_length we can lower it and reduce memory significantly, or (b) proceed directly. Which do you prefer?"

RUST_LOG=info cargo run --release

For full zero-knowledge (hides witness via BlindFold protocol), enable zk in both crates. Use jolt new --zk to scaffold a ZK project, or add manually:

Host Cargo.toml:

jolt-sdk = { git = "https://github.com/a16z/jolt", features = ["host", "zk"] }

Guest Cargo.toml:

jolt = { package = "jolt-sdk", git = "https://github.com/a16z/jolt", features = ["zk"] }

In the host, pass BlindfoldSetup to verifier preprocessing:

let blindfold_setup = prover_prep.blindfold_setup();
let verifier_prep = guest::preprocess_verifier_<fn>(shared, verifier_setup, Some(blindfold_setup));

Preprocessing runs once on first invocation and is not included in "Prover runtime". Diagnose failures:

Error	Fix
max_trace_length exceeded	Add max_trace_length = N (tight power of 2 — proving time scales with this)
heap allocation failed	Add heap_size = N
stack overflow	Increase stack_size; start at 8388608 (8 MB) if not already set
Illegal instruction	Rewrite floats as fixed-point
could not find crate	Find no_std alternative or switch to std mode
does not implement Serialize	Add #[derive(serde::Serialize, serde::Deserialize)]

Step 8 — Summarize

Tell the user: what function was made provable, what type adaptations were applied and why, std or no_std mode, and how to run it.

Once the proof runs end-to-end, always offer a performance optimization pass:

"The proof works! Want me to optimize it? I can tighten max_trace_length to reduce memory and proving time, profile which sections dominate cycle count, and offload expensive witness computation."

Step 9 — Optimize (offer after Step 8 succeeds)

Work through these in order:

1. Tighten max_trace_length — run guest::analyze_<fn>(<inputs>), find the actual cycle count, set max_trace_length to the smallest power of 2 above it. Proving time and peak memory are both proportional — a 2× reduction is a 2× speedup.

2. Find the bottleneck — add start_cycle_tracking / end_cycle_tracking (see Step 5) around major sections and run analyze_<fn> again. Focus on whichever section consumes >50% of cycles.

3. Offload expensive witness computation — if a section is expensive to compute but cheap to verify (sorting, hashing, witness generation), convert it to #[jolt::advice]. The advice function runs on the host outside the proof; the guest only verifies the result:

#[jolt::advice]
fn sort_array(input: &[u64]) -> jolt::UntrustedAdvice<Vec<u64>> {
    let mut v = input.to_vec();
    v.sort_unstable();
    v
}

#[jolt::provable]
fn my_fn(input: &[u64]) -> bool {
    let adv = sort_array(input);
    let sorted = &*adv;
    // O(n) verification: sorted order + length
    jolt::check_advice!(sorted.windows(2).all(|w| w[0] <= w[1]));
    jolt::check_advice!(sorted.len() == input.len());
    true
}

4. Use crypto inlines — for SHA-2, Keccak, secp256k1, replace standard crate calls with jolt-inlines-* (constraint-native, fraction of the cycle cost):

jolt-inlines-sha2 = { git = "https://github.com/a16z/jolt" }

5. Trim stack_size and heap_size — over-allocation doesn't cost cycles but does increase peak prover memory. Lower to actual usage once max_trace_length is tight.