Patterns from scientific Python packages

binference’s interface follows patterns that users of other inference packages praise. The census behind the field pages recorded them.

package	pattern	in binference
emcee (Foreman-Mackey et al. 2019; Foreman-Mackey et al. 2013)	plain inputs; a sampler that runs again to continue	a second engine.run(n) continues the chains without a new warm-up
scikit-learn (Buitinck et al. 2013)	hyperparameters in the constructor, fit returning self, get_params	PopulationEstimator, without a scikit-learn dependency
dynesty, nautilus	configured objects, priors declared as objects	priors and selections passed to one infer call; a response through Likelihoods.from_binned
UltraNest	output that says which criteria were met	result.check() lists failed diagnostics and remedies
ArviZ 1.x	results that other tools read	result.to_datatree() in ArviZ’s layout
SPEC 7 (van der Walt et al. 2023)	one rng argument, a seed or a Generator	a separate stream per chain; legacy RandomState rejected
tinygp	the model as functions other samplers can call	PopulationModel; PopulationResult.from_samples wraps outside draws

A scikit-learn estimator takes a feature matrix, and binference takes a likelihood matrix. Full compliance with scikit-learn therefore does not fit.

Validation follows the statistical-software literature: exact posteriors for small problems, Geweke’s “getting it right” test for Gibbs, and simulation-based calibration (Talts et al. 2018). The calibration ranks the joint log-likelihood as well as each fraction. Parameter ranks alone “could never detect large classes of problems including when the posterior is equal to the prior” (Modrák et al. 2025).