Monte Carlo

The monte_carlo module provides a bounded pipeline for repeated simulation, filtering, transformation, and diagnostic testing. The main use case is to treat one SolvedModel as the data-generating process (DGP), treat another SolvedModel as the reference model, and aggregate diagnostic test results over independent replications.

Reference and DGP Roles

The built-in simulation step draws data from dgp. The built-in filtering step then runs reference.kalman(...) on the generated observables. The reference model is not simulated by the built-in DGP pipeline.

Spec and runner exports

The serializable pipeline spec and the runner that consumes it live in the core module — no [ui] extra is required to validate, compile, or run a pipeline. The same entry points back the GUI and the .sdsge bundle.

Export	Purpose
`PipelineSpec` / `NodeSpec` / `EdgeSpec`	Pydantic-free graph specification. Serialized to JSON inside a bundle's `montecarlo/pipeline.json`.
`validate_pipeline_spec(spec, *, has_reference, has_dgp)`	Topological validation against the step-kind sets and catalog metadata; returns the ordered node list when the graph is well-formed.
`build_pipeline(ordered_nodes, *, dgp=None, resources=None)`	Compile validated nodes into an `MCPipeline` ready to run. `resources` reattaches raw-data arrays and custom callables referenced by a bundle spec.
`run_pipeline(spec, *, reference, dgp, n_rep, fail_fast, resources=None)`	One-shot validate + compile + run; returns `MCPipelineResult`.

from SymbolicDSGE import load_bundle
from SymbolicDSGE.monte_carlo import run_pipeline

loaded = load_bundle("experiment-1.sdsge")
result = run_pipeline(
    loaded.mc.spec,
    reference=loaded.reference,
    dgp=loaded.dgp,
    n_rep=500,
    fail_fast=True,
    resources=loaded.mc.resources,
)

Bundle integration

A loaded LoadedMC.spec is a PipelineSpec; LoadedMC.resources carries the side-channel arrays/callables needed by raw_data and custom nodes. See the Bundle Loading Guide for the end-to-end flow.

Step catalog

STEP_CATALOG is the registry for catalog-backed built-ins and the GUI step palette. Resource-backed node kinds such as raw_data and custom reattach large arrays or callables through the resources seam when a bundled pipeline is loaded.

Name	Purpose
`STEP_CATALOG`	Mapping from `step_type` (string) to `StepDefinition`.
`StepDefinition`	Per-step metadata: human label, parameter `FieldSpec` list, operation role, category, factory, and optional parameter compile hook.
`FieldSpec`	One parameter on a step: name, type, default, validation hints. Drives the GUI form generation.
`DATAGEN_STEP_TYPES`	Catalog-local step-kind set for valid datagen roots: `"simulation"` and `"raw_data"`.
`TRANSFORM_STEP_TYPES`	Catalog-local step-kind set for catalog-backed transforms.
`TERMINAL_STEP_TYPES`	Catalog-local step-kind set for test/regression summaries. Terminal steps cannot link forward.
`catalog_payload()`	JSON-safe rendering of the catalog for the GUI / external consumers.

Step-kind sets

The step-kind sets are implementation metadata in SymbolicDSGE.monte_carlo.catalog. They describe compiler behavior but are not the primary user-facing import surface.