Bayesian Statistician Plugin

Best practices for writing efficient, clean Stan programs. Covers structure, parameterization, prior/posterior predictive blocks, and pitfalls. Sub-references handle specialized cases.

Exploratory data analysis reference for Bayesian modeling — operational EDA procedures (data semantics audit, quality checks, timestamp handling, visualization, modeling handoff) plus a diagnostic test library indexed by data shape.

How a Stan fit becomes the canonical artifacts the workflow audits — sampling defaults, InferenceData construction, convergence and LOO numbers, the file contract — with three runnable reference scripts (posterior fit, prior predictive, fake-data simulation + recovery) to copy and adapt.

Thin phase driver for the Bayesian workflow — invokes the four bundled workflow scripts (explore, design, develop, report), holds the user gates between phases, persists the ledgers, and keeps the lab notebook. Loaded by the `/bayesian-workflow:run` command.

The writing quality bar for the final report — skim test, narrative arc, number density, figure discipline, detail tiers, and mechanical hygiene. Shared by the whole report pipeline; the critic audits against these rules by name.

Methodology for setting up a Bayesian analysis — analysis purpose, validation strategy, domain context, structural questions. Read before specifying any models.

Reference for the Bayesian workflow Python environment — uv usage, dependencies, script structure conventions, and where the Stan execution recipes live. Assumes the environment has been bootstrapped via `/bayesian-workflow:setup`.

Three-part critique methodology for a single fitted Bayesian model — statistical, domain, and framework assessment plus the verdict structure and refinement-suggestion conventions.

Style and format conventions for subagent artifacts — reports, logs, figures.

Entry/exit contract for workflow subagents — input validation, the completed-work short-circuit, structured returns with audited numeric fields, and the status.json completion record.

Methodology for fake-data simulation — drawing or fixing true parameter values, simulating data from the model, refitting, and checking that inference recovers the parameters. Covers single-draw checks (cheap) and Simulation-Based Calibration (rigorous).

ArviZ InferenceData creation from CmdStanPy — variable naming conventions, conversion arguments, common failures.

MCMC convergence diagnostics — thresholds, visual checks, and interpretation of HMC pathologies.

Required decisions for complete generative model specifications, the experiment-design discipline behind a resolution sequence, and the cross-cutting modeling principles that apply to every spec.

Comparing model populations via ELPD, decision rules for selection vs stacking, and metric validity

Guidelines for visual predictive checks following Säilynoja et al. recommendations using ArviZ

Interpreting LOO diagnostics, PIT calibration, and Pareto k for single-model evaluation