Writing a new modifier

A modifier is a wrapper type around exactly one univariate distribution that changes one behaviour and delegates the rest. This page describes the pattern the built-in modifiers follow, using them as worked examples.

The shape of a modifier

1. A wrapper struct subtyping UnivariateDistribution, holding the inner distribution plus the modifier's parameters. Validate parameters in an inner constructor — see Affine's positive-scale check in src/Affine.jl.

2. A lowercase constructor verb (affine, weight, thin, modify) as the user-facing entry point. If an "absent" parameter is meaningful, add a ::Nothing passthrough method returning the distribution unchanged, as weight(d, nothing) and thin(d, nothing) do.

3. A get_dist method exposing the inner distribution, one line:

get_dist(d::MyWrapper) = d.dist

This joins your wrapper to the unwrap protocol, so get_dist_recursive and downstream tooling work without knowing your type.

Delegation

Most methods should delegate to the inner distribution. For bulk delegation, Transformed uses an @eval loop worth copying:

for f in (:minimum, :maximum, :mean, :var, :std, :mode, :median,
    :skewness, :kurtosis, :entropy)
    @eval Distributions.$f(d::MyWrapper) = Distributions.$f(d.dist)
end

Then override only the methods your modifier actually changes (logpdf for a weight, the full change-of-variables set for an affine map). Remember Base.eltype — without it, batch sampling rand(rng, d, n) falls back to Vector{Any} and errors.

Forward-transform ops

A new forward op (alongside ThinOp/CumulativeOp) does not need a new wrapper type. Implement _apply_op(op, series) and, if the op has a displayable parameter, _op_params(op); then construct it with series_transform(d, op).

AD friendliness

Keep branches data-driven (on types and object identity), never on sampled parameter values, so gradients survive all backends. Add a gradient scenario for your modifier to test/ADFixtures/src/ADFixtures.jl, mirroring the existing entries — the AD CI sweeps it across ForwardDiff, ReverseDiff, Enzyme and Mooncake automatically.

Checklist

• [ ] Struct + validated inner constructor

• [ ] Constructor verb (+ Nothing passthrough if applicable)

• [ ] get_dist method

• [ ] Delegations + overridden behaviour

• [ ] Base.eltype

• [ ] Docstrings in the house style (@doc blocks, # Examples, # See also)

• [ ] Export the verb; mark the type public in src/public.jl

• [ ] Tests in test/MyWrapper.jl + an ADFixtures scenario