MCPipeline

class MCPipeline(steps: Sequence[MCStep])

MCPipeline stores the ordered operations executed inside each Monte Carlo replication.

Contract:

Rule	Description
One data-generation step	The first step must have `op_type=OpType.DATAGEN`. Later steps cannot generate data once `DATAGEN` is performed.
Unique step names	Names are used as payload and result keys.
Per-replication payloads	Steps may pass results through `MCContext.payloads` inside the same replication. Aggregate summaries are created after all successful replications finish.

Methods:

MCPipeline.graph -> PipelineGraph

Return the cached dependency graph inferred from the pipeline's step kwargs. The graph records structural edges used by serialization, including filter dependencies and payload-producing transform/custom steps.

MCPipeline.to_spec() -> PipelineSpec

Serialize the live pipeline into the graph-form PipelineSpec. Bulk side channels are referenced by key: raw_data arrays and custom callables are written as separate bundle members by BundleBuilder.add_mc(...).

MCPipeline.run(
    *,
    reference: SolvedModel,
    dgp: SolvedModel | None = None,
    n_rep: int,
    retain_payloads: bool = True,
    retain_test_results: bool = True,
    retain_contexts: bool = False,
    fail_fast: bool = True,
    verbosity: int = 1,
) -> MCPipelineResult

Inputs:

Name	Description
reference	Reference `SolvedModel` used by reference-side operations such as Kalman filtering.
dgp	Optional DGP `SolvedModel`. Required by `simulation_step`, but not required by `raw_data_step`.
n_rep	Number of Monte Carlo replications.
retain_payloads	Store each successful replication's payload dictionary in the result container.
retain_test_results	Store scalar `TestResult` objects from each successful replication.
retain_contexts	Store full `MCContext` objects. This is useful for debugging and memory-heavy for large runs.
fail_fast	If `True`, raise on the first failed replication. If `False`, collect `MCFailure` entries and summarize successful replications.
verbosity	Performance-reporting level: `0` prints nothing, `1` prints one aggregate throughput line, and `2` prints one throughput line per step.

Returns:

Type	Description
`MCPipelineResult`	Aggregate container with test summaries, optional per-replication payloads, optional scalar test results, optional contexts, and failures.

Serializable steps

to_spec() requires each step to carry a step_type. Use the built-in factories rather than hand-building MCStep objects when the pipeline needs to enter a .sdsge bundle.