Concepts and reading paths

Foerster runs probabilistic programs on Spindel's forkable execution runtime. These five objects have different roles and lifetimes:

Object	Meaning	Boundary
World	An execution context containing managed application state and registered systems. Canonical particle methods copy the caller's world; fresh-world methods start without it.	Forking isolates state supported by the runtime or a system adapter. Ordinary host atoms and arbitrary I/O are outside that boundary.
Trace	One run's addressed random choices, observations, scores and replay information. Trace operations can propose changes and re-execute a model.	Live replay resources have ownership and cleanup obligations. An arbitrary collection of values is not a replayable trace.
Measure	A settled inference result, with weighted values, summaries and method-specific diagnostics.	It is not a collection of live application worlds. Normalizer/evidence availability depends on the estimator.
Population	Intermediate inference state over a target, with weights and retained trace or numeric values. Population programs compose target changes, resampling and moves.	A move's invariance, a proposal correction and an evidence estimate are separate contracts. A common result interface does not certify all three.
Numerical block	A site exposing a joint numeric state and explicit density, gradient or other sampler capabilities. Raster can compile those supplied functions.	Foerster does not automatically extract a differentiable density from an arbitrary simulator.

A typical particle run starts from a world, records traces, maintains a population, then settles a measure and retires its owned worlds. Numerical methods can instead operate on block values. Read the algorithm guide for each entry point's model and world-policy requirements, and the programming model for how proposals, kernels, pairs and population programs plug into each other.

Choose a path

• First probabilistic program: Getting Started, then Choosing an Inference Algorithm and checking a model.
• Reuse Clojure application code: Inference in a Running Application, then Models in Worlds and the language scope rules.
• Program inference: Inference Programs for weighted populations; Trace Programming and Custom Moves for the generative function interface and custom trace proposals.
• Use numerical inference: Blocks and HMC, then From the PyMC Gallery. foerster-raster supplies compiled block functions and automatic differentiation.

Causal inference, amortized inference and incremental trace inference each have additional model assumptions and current restrictions. In particular, counterfactuals need structural mechanisms and a declared noise coupling; an amortized sampler is an approximation whose training loss alone does not validate its posterior; fixed-target trace reuse does not implement a general reactive posterior controller.