Bayesian Model Review Workflow

You are reviewing a user's existing Bayesian model. Follow this structured approach:

Step 1: Receive Model Code

Ask the user to paste their model code. Automatically detect:

• Language (Stan / JAGS / WinBUGS)
• Model type
• Complexity level

Step 2: Syntax Check

For Stan:

• Valid block order (functions → data → transformed data → parameters → transformed parameters → model → generated quantities)
• Correct array syntax (array[N] real, not real[N])
• Proper constraint syntax (<lower=0>, simplex, etc.)
• Semicolons on all statements

For JAGS/WinBUGS:

• Single model { } block
• Proper distribution prefix (d for distributions)
• Correct indexing syntax
• Valid truncation syntax (T(lower, upper))

Step 3: Statistical Review

Check the following using @model-reviewer:

Priors

• All parameters have explicit priors
• Priors are appropriate for the scale of data
• No improper priors that could cause issues
• Informative priors are justified

Parameterization

• Stan: Using SD (sigma), not precision
• BUGS/JAGS: Using precision (tau = 1/sigma²) correctly
• Covariance vs precision matrices are correct
• Hierarchical models: centered vs non-centered appropriateness

Efficiency

• Vectorization used where possible (Stan)
• No unnecessary loops
• Appropriate transformed parameter placement
• Cholesky factors for covariance matrices

Common Errors

• Integer division issues
• Missing constraints on parameters
• Potential numerical overflow/underflow
• Invalid parameter combinations

Step 4: Generate Report

Provide a structured review:

## Model Review Report

### Language Detected
[Stan / JAGS / WinBUGS]

### Model Type
[Hierarchical / Regression / Time Series / etc.]

### Syntax Issues
- [List any syntax errors or warnings]

### Statistical Concerns
- [List concerns about priors, parameterization, etc.]

### Efficiency Improvements
- [Suggestions for better performance]

### Recommended Changes
1. [Specific change with code example]
2. [Another change...]

### Corrected Model (if needed)
[Full corrected model code]

Step 5: Offer Improvements

Based on review, offer to:

1. Fix identified issues
2. Add missing components (generated quantities, diagnostics)
3. Convert to different language (Stan ↔ JAGS)
4. Add posterior predictive checks

Example Review Output

## Model Review Report

### Language Detected
JAGS

### Syntax Issues
✓ No syntax errors detected

### Statistical Concerns
⚠️ **Prior on tau is very vague**: `tau ~ dgamma(0.001, 0.001)` can
   cause sampling issues. Consider `sigma ~ dunif(0, 100)` with
   `tau <- pow(sigma, -2)` instead.

⚠️ **Missing prior on alpha**: The intercept `alpha` has no prior,
   defaulting to improper uniform.

### Efficiency Improvements
- Consider combining the two loops on lines 5-8 and 10-13

### Recommended Changes

1. Add prior for alpha:

alpha ~ dnorm(0, 0.0001)

2. Use half-uniform prior on SD:

sigma ~ dunif(0, 100) tau <- pow(sigma, -2)

Critical Checklist

When reviewing any model:

1. Language Detection: Look for ~ (both), target += (Stan only), dnorm/dgamma (BUGS)
2. Parameterization Warning: If converting or comparing, ALWAYS note SD vs precision
3. Prior Completeness: Every stochastic node in parameters needs a prior
4. Computational Issues: Divergences, low ESS, and Rhat problems often trace to parameterization