Skip to content
Aleatoric Systems

Determinism

What this means in practice

Section titled “What this means in practice”

Given the same SimulationManifest (including the same seed), you get the same event sequence whether you:

  • generate a batch dataset, or
  • consume a stream of the same simulation.

This is the foundation for reproducible experiments.

Determinism pipeline: same manifest and seed produce identical batch and stream outputs.

The Guarantee

Section titled “The Guarantee”

Given:

  • Config C (with Seed S)
  • Generation Function G(C)

Then: Events_Batch = G_batch(C) Events_Stream = G_stream(C)

Assert: Events_Batch == Events_Stream

This property is useful when you want a model, strategy, or pipeline to see the same conditions across environments.

Implementation Details

Section titled “Implementation Details”

Seed Management

Section titled “Seed Management”

The HyperSynthReactor initializes its internal random number generators (NumPy RandomState or Python random) using the provided seed at instantiation.

Shared Logic

Section titled “Shared Logic”

Both Batch and Stream drivers utilize the same underlying generator instance (market.stream()).

  • Batch consumes the generator greedily.
  • Stream consumes the generator lazily, inserting asyncio.sleep() calls between yields.

Because the underlying math (GBM steps, orderbook logic) happens inside market.stream() before the driver decides to sleep or append, the sequence of internal states remains identical.

Floating Point Consistency

Section titled “Floating Point Consistency”

We prioritize integer-based timestamps (timestamp_ms) and consistent floating-point operations within the single Python process to avoid non-deterministic behavior often seen in multi-threaded or distributed simulations.

Verification

Section titled “Verification”

The guarantee is enforced by the test suite in tests/test_determinism.py.

async def test_batch_stream_parity():
    # 1. Generate Batch
    batch_file = await unified_driver(config, mode="batch", ...)
    batch_events = read_parquet(batch_file)


    # 2. Generate Stream
    stream_gen = await unified_driver(config, mode="stream", ...)
    stream_events = [json.loads(e) async for e in stream_gen]


    # 3. Assert Equality
    assert len(batch_events) == len(stream_events)
    for b, s in zip(batch_events, stream_events):
        assert b == s

This test is part of the CI/CD pipeline and must pass for any release.

What’s next

Section titled “What’s next”