Configuration

A prover's behavior is set by one StateSyncGkrConfig value. Its fields are plain values. Changing smt or layer_strategy changes the compiled circuit identity, its commitment and every proof made under it; batching only affects callers that drive batching::BatchProver. This page lists each value with its default and limits, then the parameters that are fixed in the source.

StateSyncGkrConfig#

use statesync_gkr::StateSyncGkrConfig;
use statesync_gkr::batching::BatchPolicy;
use statesync_gkr::compiler::{LayerStrategy, SmtParams};

fn main() {
    let config = StateSyncGkrConfig {
        smt: SmtParams {
            depth: 24,
            leaf_max_fields: 31,
        },
        layer_strategy: LayerStrategy::A,
        batching: BatchPolicy::default(),
    };
    // The explicit values above are the defaults.
    assert_eq!(config.smt, SmtParams::default());
    assert_eq!(config.layer_strategy, LayerStrategy::default());
    println!("{config:?}");
}

Build PreparedSync after you fix the configuration, because it encodes that configuration's circuit. Do not change prover.config while you reuse prepared state; build a new preparation instead.

Tree depth#

SmtParams::depth is the number of tree levels, which is also the number of siblings in every path.

A larger depth widens the circuit and increases witness, proving and verification work. In the measured profiles the layer count stayed at 118 and the encoded membership proof grew about 2.93% from depth 24 to depth 32; Circuits and layers explains why. Each depth compiles to a different circuit with a different commitment and prepared state, so trees of different depths are separate instances.

Leaf bound#

SmtParams::leaf_max_fields is the maximum length of a leaf encoding in field elements, tag included.

The bound is part of the circuit identity and the circuit commitment, so any change produces a different circuit identity and commitment. A change that alters the pre-image width also changes every leaf digest, and with it every root, so treat the bound like the depth: fixed for the life of a tree. The 118-layer schedule was measured at the default bound; a bound with a different pre-image width changes the leaf hash circuit, and that measurement no longer applies.

Layer strategy#

LayerStrategy selects how tree levels map to GKR layers.

The unsupported variants fail with the reason "v0.1 implements strategy A (one data-parallel instance per tree level) only". The inner-proof encoding accepts only the strategy identifier of A.

Batch policy#

batching::BatchPolicy holds the knobs of the in-process deadline scheduler.

DeadlineScheduler::decide_batch_size(queue_len, remaining) returns 0 for an empty queue and the whole queue when it is shorter than min_batch_size. Otherwise it takes up to max_batch_size jobs, and when remaining is below deadline it scales that size down in proportion, never below min_batch_size.

The defaults are provisional values for this scheduler. The measured serving settings (worker count, batch cap and maximum wait) belong to a separate benchmark caller and are covered in Performance tuning.

Fixed parameters#

These values are compiled in. There is no runtime API to select a different field, hash or transcript.

Changing the cryptographic profile#

The generic traits, HashGadget, WiringOracle, SumcheckOracle and the Field bounds of the sumcheck and circuit types, are extension points. The shipped proving and verification functions, however, use the fixed aliases above. Replacing the field, the hash, the transcript or the compiler is a new implementation and a new verification task:

• proof bytes, circuit commitments and inner-proof compatibility change, and the transcript change requires a new protocol_version;
• the existing proof evidence and formal results do not transfer automatically.

Custom frontends describes what a new frontend over the existing field and hash must supply.

Limits at a glance#

Errors describes every error type and how to handle it.