Interface overview

The public surface of Reactant.ProbProg is partitioned into four groups: modeling primitives, inference drivers, persistent state, and data types. The full signatures and semantics are documented in the API reference; the tables below are a quick index.

Modeling primitives

Symbol	Signature	Role
sample	sample(rng, dist; symbol)  /  sample(rng, f, args...; symbol, logpdf, support, bounds)	Record a random choice from a distribution or user-defined sampler.
untraced_call	untraced_call(rng, f, args...)	Execute a probabilistic function without recording its choices.
simulate  /  simulate_	simulate(rng, f, args...)	Forward-simulate a model; return its trace and prior log-density.
generate  /  generate_	generate(rng, constraint, f, args...)	Execute a model conditioned on observations; return the trace and log importance weight.

Inference

Symbol	Signature	Role
mh	mh(rng, trace, weight, f, args...; selection)	One Metropolis-Hastings step regenerating the selected addresses.
mcmc	mcmc(rng, trace, f, args...; selection, algorithm, num_samples, ...)	Trace-based HMC or NUTS.
mcmc_logpdf	mcmc_logpdf(rng, logdensity_fn, initial_position, args...; algorithm, ...)	HMC or NUTS over a user-supplied log-density.
run_chain	run_chain(rng, logpdf_fn, initial_position, args...; num_warmup, num_samples, chunk_size, ...)	Chunked chain driver with progress bar and resumable state.

State and persistence

Symbol	Signature	Role
MCMCState	(struct) position, gradient, potential_energy, step_size, inverse_mass_matrix, rng	Resume token returned by every inference entry point.
save_state	save_state(filename, state)	Serialise an MCMCState to disk.
load_state	load_state(filename)	Reload an MCMCState from disk.

Data types and distributions

Symbol	Role
Trace	Executed-model record: choices, subtraces, retval, weight.
Constraint	Dict-like Address → value mapping for observations.
Address	Immutable symbol path identifying a sample site.
Selection	Ordered set of addresses; constructed with select.
Normal  /  Exponential  /  LogNormal  /  Bernoulli	Built-in distributions, each carrying a static shape::Tuple and registered sampler / log-density.