What problem does it solve? Writing PyMC models that sample without errors is not the same as writing models that are statistically correct. This Skill guides the construction of identifiable generative models with named coordinates and dims, enforces prior predictive and simulated-recovery testing before fitting, and treats convergence diagnostics and posterior predictive checks as mandatory correctness gates rather than optional extras. ## Core Features & Use Cases - Model construction with shape contracts: Build pymc.Model graphs with coords/dims, Data containers, random variables, deterministics, and potentials whose shapes and coordinate lengths are asserted before sampling. - Inference and diagnostics workflow: Run multi-chain sampling with recorded seeds, then gate on divergences, R-hat, ESS, MCSE, energy/BFMI, and tree depth instead of trusting a single trace plot. - Prediction and comparison recipes: Update Data and coordinates together for out-of-sample prediction, keep predictions separate from in-sample posterior predictive draws, and attach pointwise log likelihood for ArviZ model comparison. - Use Case: You are fitting a hierarchical logistic regression and hitting divergences. The Skill walks you through checking the funnel geometry, switching to a non-centered parameterization, validating prior predictive plausibility, and confirming convergence across four chains before trusting the posterior. ## Quick Start Use the pymc-python skill to build and diagnose a hierarchical Bayesian regression model with prior predictive checks and out-of-sample prediction on new data.
