zk-scribble

Trace mutation fuzzer for Hekate ZK programs and chiplets.

Tampers your valid trace, runs preflight checks, panics if the tamper goes undetected. If scribble panics, your constraints have a hole.

What it does

Flips bits, swaps rows, toggles selectors, injects out-of-range values
Validates using preflight (row-by-row constraint eval + bus multiset check)
Proptest integration, shrinks failures to the smallest mutation that escapes your AIR
Runs in debug mode in seconds, not minutes

What it doesn't do

No proofs. Scribble never calls the prover or verifier. It checks your constraints on the concrete trace, not through the ZK pipeline. This is why it's fast.
No protocol-level testing. Transcript binding, Fiat-Shamir, evaluation arguments, Brakedown, those need the real prover/verifier e2e tests.
No soundness guarantees. Passing scribble means every random mutation was caught. It doesn't prove your constraints are complete, only that the ones you wrote are wired up.

Two layers

Layer 1: random coverage. assert_all_caught with shallow mutations (BitFlip, FlipSelector, SwapRows, OutOfBounds). Finds unconstrained columns, missing booleans, selector isolation gaps. Proptest shrinks any escape to a minimal reproduction. Run during chiplet development.

Layer 2: deterministic emulation. check_single_mutation with hand-crafted structural or compound mutations (SwapColumns, CopyColumns, Compound). Replaces the 80-line boilerplate of each e2e exploit test with a 15-line call. Same attack logic, 100x faster (preflight vs prove+verify). Run as regression suite.

Quick start

use zk_scribble::{ScribbleConfig, assert_all_caught};

#[test]
fn my_chiplet_survives_chaos() {
    let (air, instance, witness) = setup_my_chiplet();
    assert_all_caught(&air, &instance, &witness, ScribbleConfig::default());
}

Target a specific chiplet

use zk_scribble::{ScribbleConfig, Target, assert_all_caught};

#[test]
fn mlkem_ntt_chiplet_survives_chaos() {
    let (program, instance, witness) = setup_mlkem_fixture();

    let config = ScribbleConfig::default()
        .target(Target::Chiplet(1))
        .cases(512);

    assert_all_caught(&program, &instance, &witness, config);
}

Restrict mutation kinds

use zk_scribble::{MutationKind, ScribbleConfig, assert_all_caught};

#[test]
fn ram_selector_fuzzing() {
    let (air, instance, witness) = setup_ram_fixture();

    let config = ScribbleConfig::default()
        .mutations([MutationKind::FlipSelector, MutationKind::SwapRows])
        .cases(1024);

    assert_all_caught(&air, &instance, &witness, config);
}

Dispatch swap (Layer 2)

Swap a subset of columns between two rows. Selectors and RAM columns stay intact, emulates an attacker who rearranges data while preserving dispatch structure.

use zk_scribble::{Mutation, Target, check_single_mutation};

#[test]
fn ntt_dispatch_swap_caught() {
    let (air, instance, witness) = setup_ntt_fixture();

    let ntt_data_cols = vec![
        NTT_A, NTT_B, NTT_A_OUT, NTT_B_OUT,
        NTT_LAYER, NTT_BFLY, NTT_INSTANCE,
    ];

    let mutation = Mutation::SwapColumns {
        target: Target::Chiplet(0),
        cols: ntt_data_cols,
        row_a: 5,
        row_b: 12,
    };

    let result = check_single_mutation(&air, &instance, &witness, &mutation);
    assert!(result.is_ok(), "dispatch swap must be caught by RAM binding");
}

Coordinated cross-trace attack

Compound applies multiple mutations atomically. Modify chiplet AND main trace in sync to bypass per-table checks, the cross-table bus is the only backstop.

use zk_scribble::{Mutation, Target, check_single_mutation};

#[test]
fn out_of_range_cross_trace_caught() {
    let (program, instance, witness) = setup_ntt_with_cpu_fixture();

    let mutation = Mutation::Compound(vec![
        Mutation::OutOfBounds {
            target: Target::Chiplet(0),
            col: BUS_B_OUT_PHY,
            row: 0,
            value: Q as u128,
        },
        Mutation::OutOfBounds {
            target: Target::Main,
            col: CPU_B_OUT,
            row: 0,
            value: Q as u128,
        },
    ]);

    let result = check_single_mutation(&program, &instance, &witness, &mutation);
    assert!(result.is_ok(), "out-of-range b_out must be caught");
}

Reading the preflight report

When preflight catches a mutation, the report names the failing invariant, table, and row:

  PREFLIGHT: 1 constraint violations, 0 boundary violations, 0 bus issues

    [Chiplet 0] Constraint 32 "bits_01_equal" failed at row 0

Boundary violations show the concrete mismatch:

  PREFLIGHT: 0 constraint violations, 2 boundary violations, 0 bus issues

    Boundary #0: col=0 row=0 actual=Flat(Block128(0)) expected=Flat(Block128(22067681354706156661646625971774519825))
    Boundary #1: col=1 row=0 actual=Flat(Block128(1)) expected=Flat(Block128(265498766201044366875656389800751278795))

Bus diagnostics list every endpoint with its row count, active rows, and multiset product, so you can see which side of the bus diverged:

  PREFLIGHT: 0 constraint violations, 0 boundary violations, 1 bus issues

    Bus "test_bus" (2 endpoints):
      Main: 8 rows, 4 active, product=Flat(Block128(16200159481073039905153729824056248498))
      Chiplet 0: 8 rows, 4 active, product=Flat(Block128(179105452667142969769264969604701840821))

If a mutation escapes (report is clean), scribble panics with the proptest-shrunk Mutation. That tamper is the soundness gap, add a constraint that catches it.

Requirements

Depends on hekate-math, hekate-sdk, hekate-core, hekate-program.

License

MIT

zk-scribble 0.1.0