baywright

Use this agent to interpret the output of MCMC and calibration diagnostics — R-hat, ESS, divergences, E-BFMI, trace/rank plots, SBC rank histograms, LOO-PIT, Pareto k — in plain language and recommend the next action. Invoke when someone pastes or describes diagnostic output and needs to know what it means and what to do. <example>Context: The user ran a sampler and got warnings. user: "I got 47 divergences and a bulk-ESS of 80, what does that mean?" assistant: "Let me hand this to the diagnostic-reader agent to interpret and recommend the fix." <commentary>Translating raw diagnostics into a verdict and a concrete next action is this agent's specialty.</commentary></example> <example>Context: The user shows an SBC rank histogram. user: "The SBC ranks pile up at both ends — is that bad?" assistant: "Let me use the diagnostic-reader agent to read that histogram." <commentary>A U-shaped SBC histogram is overconfidence; the agent names it and prescribes the honest fix.</commentary></example>

Use this agent to adversarially criticize a Bayesian model — to find how it could be wrong and what would falsify it — by proposing posterior-predictive test quantities and identifiability / prior-sensitivity probes. Invoke during the criticism stage of the baywright workflow, or whenever someone is ready to trust or ship a model. <example>Context: A model has converged and the user is satisfied. user: "The model fits great, I think we're done." assistant: "Before we call it done, let me run the model-critic agent to try to break it." <commentary>The critic hunts for the misfit the user is not looking for — exactly when someone is ready to stop is when it is most useful.</commentary></example> <example>Context: The user reports a clean posterior. user: "Everything looks good, R-hat is 1.0." assistant: "Clean sampling isn't adequacy — let me use the model-critic agent to design posterior-predictive checks." <commentary>The agent distinguishes 'the sampler worked' from 'the model is right' and proposes the checks that tell them apart.</commentary></example>

Use this agent to elicit priors for a Bayesian model through a Socratic interview — turning a domain expert's vague beliefs into prior distributions on the right scale and pressure-testing them against the prior predictive. Invoke during the priors stage of the baywright workflow, or whenever someone needs help turning "I'm not sure what prior to use" into a defensible choice. <example>Context: The user is choosing priors for a hierarchical model. user: "I have no idea what prior to put on the group-level scale." assistant: "Let me bring in the prior-interviewer agent to elicit it with you." <commentary>Prior elicitation is precisely this agent's job: interview about plausible magnitudes, propose a prior on the right scale, and name the prior-predictive check that confirms it.</commentary></example> <example>Context: A user has written flat priors everywhere. user: "I just used uniform priors so it's objective." assistant: "Let me use the prior-interviewer agent to check what those priors actually imply on the outcome scale." <commentary>The agent's value here is exposing that 'uninformative' priors are often strongly informative on the outcome scale.</commentary></example>

The entry point and orchestrator for Bayesian modeling: load this for any end-to-end modeling effort, any task that spans more than one stage, or the meta-question "is my model any good / good enough / done?". Runs an interactive "super-REPL" — the agent drives a live, hot-reloading session, interprets what it prints, and guides the human at every turn — and enforces a fixed sequence (formulate -> priors -> fit + diagnose -> calibrate -> criticize -> compare -> report) plus non-negotiable honesty gates (trust calibration first; never tune-to-pass; never call a model "good" without evidence). Methodology is prose + math only, tool-agnostic (Julia-first via Revise.jl; also Stan/PyMC/Turing/NumPyro/brms/R); no baked code — consult current docs and write live code in the session. The individual stages each have their own skill for narrow, stage-specific questions; this one routes and sequences them. Trigger on: build or critique a Bayesian/probabilistic model, set up a Bayesian workflow, what order the workflow steps go in, or "is my model any good / good enough / done?".

Baywright workflow verb (invoke as /baywright:bw-criticize): run a posterior-predictive model criticism pass — try to break the model with test quantities, check identifiability and prior sensitivity — and record what you found in the Model Ledger. Trigger on: /baywright:bw-criticize, "criticize my model", "run posterior predictive checks", "can my model reproduce the data".

Baywright teaching verb (invoke as /baywright:bw-explain <concept>): explain a Bayesian-workflow concept in plain language — what it is, why it matters, how to read it — then point to the current docs for the chosen tool instead of pasting code. For learners. Trigger on: /baywright:bw-explain, "explain <concept>", "what is SBC / R-hat / ELPD / a funnel", "help me understand this diagnostic".

Baywright workflow verb (invoke as /baywright:bw-gate): run the calibration gate — prior predictive sanity, simulation-based calibration (SBC), and parameter recovery — fail-closed, and record the verdict in the Model Ledger. Trigger on: /baywright:bw-gate, "run the calibration gate", "check calibration", "run SBC", "is my model calibrated".

Baywright workflow verb (invoke as /baywright:bw-next): advance to the next workflow stage, but gate-check the current one first — it refuses to advance past a stage whose evidence isn't recorded. Trigger on: /baywright:bw-next, "next step in the workflow", "advance the model", "what do I do next with baywright".

1 hook across 1 event