Project Developer Guide

These are the established conventions and practices for this codebase.

1. Library Source is Local — Look It Up

All PureScript dependencies are installed with full source code in .spago/p/. When uncertain about a function's behavior, signature, or available instances — look it up.

.spago/p/prelude-7.0.0/src/...
.spago/p/transformers-6.1.0/src/...
.spago/p/foldable-traversable-7.0.0/src/...

Use whatever tool fits the situation:

• •repomix to understand a whole library's structure
• •grep to find a specific function or type
• •read when you know exactly where to look

Do not guess at function signatures or behaviors. Do not rely on potentially stale knowledge. The source is right there.

2. Property-Based Testing Only

All testing uses spec, quickcheck, and spec-quickcheck. We do not write one-off unit tests or example-based tests.

Tests assert properties over randomly generated inputs. This is appropriate for the domain — cryptographic primitives and circuit correctness are about invariants holding universally, not about specific examples passing.

Consult .spago/p/ for these libraries when needed:

• •spec — test organization and assertions
• •quickcheck — property-based testing, Arbitrary instances
• •spec-quickcheck — integration between the two

3. Circuit Testing via snarky-test-utils

For testing arithmetic circuits, use packages/snarky-test-utils. Do not invent new testing infrastructure.

The core pattern is CircuitSpec:

type CircuitSpec f c r m a avar b =
  { builtState :: CircuitBuilderState c r
  , solver :: SolverT f c m a b
  , checker :: Checker f c
  , testFunction :: a -> Expectation b    -- pure reference function
  , postCondition :: PostCondition f c r
  }

The test framework:

1. •Runs the circuit solver to produce output
2. •Runs the pure testFunction to get expected output
3. •Verifies all constraints are satisfied
4. •Compares: circuit output must equal pure function output

The utilities in snarky-test-utils are used throughout the codebase. If you need a testing pattern, it's almost certainly already there.

4. FFI: The Three-Layer Pattern

Cryptographic primitives come from Rust via packages/crypto-provider (the snarky-crypto node module). The FFI follows a three-layer architecture:

Layer 1: Rust (via napi-rs)

• •Exposes minimal surface from arkworks / o1labs/proof-systems
• •Functions take simple arguments (primitives, arrays) — not complex structs
• •Sharing Rust objects across FFI is painful, so we avoid it

Layer 2: JavaScript Wrapper

• •Each PureScript module with foreign import has a corresponding .js file
• •Imports from snarky-crypto
• •Handles translation: restructures flat arrays into records, applies currying, shapes data for PureScript
• •Absorbs awkwardness so both Rust and PureScript have clean interfaces

Parse at the FFI Boundary, Not in PureScript

When FFI returns flat/unstructured data (like [x0, y0, x1, y1, ...]), transform it in the JS layer into properly structured records that match domain semantics.

Why:

1. •PureScript only sees clean, typed structures — Vector 16 (LrPair f) instead of Array f
2. •Types are self-documenting — { l :: AffinePoint f, r :: AffinePoint f } immediately conveys meaning
3. •Parsing happens once at the boundary — not scattered throughout PureScript code
4. •No unsafe array indexing in PureScript — the JS layer handles flat-to-structured conversion

Example: Rust returns lr pairs as flat coordinates [l0.x, l0.y, r0.x, r0.y, l1.x, ...]

// In .js file — parse into structured records
export const proofOpeningLr = (proof) => {
  const flat = snarky.proofOpeningLrFlat(proof);
  const pairs = [];
  for (let i = 0; i < flat.length; i += 4) {
    pairs.push({
      l: { x: flat[i], y: flat[i + 1] },
      r: { x: flat[i + 2], y: flat[i + 3] }
    });
  }
  return pairs;
};

-- In .purs file — clean typed interface
foreign import proofOpeningLr :: Proof g f -> Array { l :: AffinePoint f, r :: AffinePoint f }

The JS layer acts as a marshalling layer between native representation and the PureScript type system.

Layer 3: PureScript

• •foreign import declarations with proper types
• •foreign import data for opaque Rust objects
• •Type classes to abstract over variants (e.g., Pallas/Vesta curves)
• •Functional dependencies for type inference

Example: See packages/pickles/test/Test/Pickles/ProofFFI.{purs,js} for a comprehensive example of this pattern.

5. Circuit vs Pure Functions — Rust is Ground Truth

Many operations exist in two forms:

• •Circuit version — uses the snarky DSL, creates constraints
• •Pure version — regular computation, no constraints

Where pure functions come from

Pure functions are typically thin wrappers around Rust FFI, not PureScript implementations. The Rust code (especially o1labs/proof-systems) provides the ground truth.

Why? This is cryptography. Complex arithmetic sequences must match protocols exactly. Since we're defining circuits that compute these protocols, we need "one foot in truth" — the pure reference function must come from a known-correct implementation.

Where pure functions live

• •In library code when needed at circuit boundaries (constructing inputs/outputs) or reused across packages
• •In tests when only needed for testing a specific circuit

Testing enforces correctness

The snarky-test-utils pattern ensures circuits match their pure counterparts:

Circuit output == Pure function output (backed by Rust)

If these disagree, the test fails. This is how we know circuits are correct.

6. The Advice Pattern — Witness Data via the m Parameter

The Snarky c t m monad has three type parameters:

• •c — constraint type (e.g., KimchiConstraint f)
• •t — state type for circuit building
• •m — the advice monad for providing witness data

What is advice?

Circuits sometimes need data that can't be computed from circuit variables alone. For example:

• •Looking up an account address from a public key
• •Factoring a number
• •Fetching a Merkle path from a tree

This data must be "conjured up" by the prover during witness generation. The m parameter is how we abstract over this.

The pattern

1. Define a typeclass for your advice:

-- Simple example from snarky-test-utils/src/Test/Snarky/Circuit/Factors.purs
class Monad m <= FactorM f m where
  factor :: F f -> m { a :: F f, b :: F f }

2. Use it in your circuit via exists:

factorsCircuit :: forall t m f c. FactorM f m => CircuitM f c t m => FVar f -> Snarky c t m Unit
factorsCircuit n = do
  { a, b } <- exists do
    nVal <- read n          -- read the circuit variable's concrete value
    lift $ factor @f nVal   -- call into the advice monad
  -- Now a and b are circuit variables (FVar f)
  -- The prover provided their values, the circuit constrains them
  c1 <- equals_ n =<< mul_ a b
  assert_ c1

3. Provide instances for different phases:

-- For proving: actually compute the witness
instance PrimeField f => FactorM f Gen where
  factor n = do
    a <- arbitrary `suchThat` \a -> a /= one && a /= n
    pure { a, b: n / a }

-- For compilation: crash (should never be called)
instance FactorM f Effect where
  factor _ = throw "unhandled request: Factor"

Two monads, two phases

Testing with advice

Use circuitSpec' (not circuitSpecPure') when your circuit needs advice:

circuitSpec' 100 randomSampleOne  -- randomSampleOne :: Gen ~> Effect
  { builtState: s
  , checker: eval
  , solver: solver
  , testFunction: satisfied_
  , postCondition: postCondition
  }
  gen

The second argument is a natural transformation m ~> Effect that runs the advice monad.

Full example

See packages/example/ for a complete worked example:

• •src/Snarky/Example/Circuits.purs — circuits using AccountMapM and MerkleRequestM
• •test/Test/Snarky/Example/Monad.purs — TransferRefM (proving) and TransferCompileM (compilation) instances
• •test/Test/Snarky/Example/Circuits.purs — testing with circuitSpec'

When to use advice

Use advice when your circuit needs witness data that:

• •Comes from external state (databases, trees, maps)
• •Requires non-trivial computation (factoring, searching)
• •Depends on prover-only knowledge

Do not use advice for values that can be computed purely from circuit inputs — just compute them directly.

7. Prefer Statically Sized Types

In the context of circuits, proofs, and cryptographic verification, data is almost always statically sized. Use PureScript's type-level sized containers instead of dynamic collections.

The types

Why this matters

1. •Circuit data is fixed-size — IPA rounds, witness columns, polynomial evaluations all have known sizes at compile time
2. •Type safety catches bugs — mismatched vector lengths are compile errors, not runtime surprises
3. •Zero runtime cost — PureScript newtypes compile away; Vector n a is just Array a at runtime

FFI boundary

Even when the underlying JavaScript/Rust type is dynamically sized (e.g., Array), use static types on the PureScript side:

-- The JS returns an Array, but we know it's always 15 elements
foreign import proofWitnessEvals :: Proof g f -> Vector 15 (PointEval f)

-- The challenges are always IPA_ROUNDS long (e.g., 16 for SRS size 2^16)
foreign import proofBulletproofChallenges :: ... -> Array f  -- Convert to Vector d f

The JS wrapper can return a plain array; the PureScript foreign import declares the static type. If sizes don't match, you'll get a runtime error — which is the correct behavior (it indicates a bug in the Rust/JS layer or a misunderstanding of the protocol).

When to determine sizes

Before writing code, investigate:

1. •What is this data? (e.g., "bulletproof challenges")
2. •What determines its size? (e.g., "IPA rounds = SRS log2 size")
3. •Is it fixed for a given circuit/proof? (almost always yes)

Then use the appropriate type-level natural.

Example: Converting at boundaries

-- FFI returns Array (dynamic)
challengesArray :: Array f
challengesArray = proofBulletproofChallenges proverIndex { proof, publicInput }

-- Convert to Vector (static) — fails if length doesn't match
challenges :: Vector 16 f
challenges = unsafePartial $ fromJust $ Vector.toVector challengesArray

The unsafePartial is acceptable here because a length mismatch indicates a protocol violation, not normal program flow.

Summary