Migration: declare predictors through the family

SuperLSS now takes the family as its first positional argument, followed by one declaration for every parameter. Terms include their input-column names. This replaces the family= and predictors= keyword constructor.

Replace predictor templates with family helpers

Before:

from superglm import GaussianLS, Numeric, Predictor, Spline, SuperLSS

family = GaussianLS()
model = SuperLSS(
    family=family,
    predictors=(
        Predictor("location", {"age": Spline(kind="cr", k=8), "value": Numeric()}),
        Predictor("scale", {}),
    ),
)

After:

from superglm import GaussianLS, SuperLSS, s

family = GaussianLS()
model = SuperLSS(
    family,
    family.location(s("age", kind="cr", k=8), "value"),
    family.scale(),
)

An empty helper replaces the empty feature mapping for an intercept-only predictor. A bare column name replaces Numeric(). Use cat("region") for a categorical effect, re("broker") for a random effect, or term("column", existing_spec) for another feature specification.

The old constructor keywords raise TypeError. Lower-level Predictor templates remain available for inspecting model configuration and for numerical interfaces; the public constructor takes family-bound declarations.

Unpack a sequence when building declarations programmatically

family = GaussianLS()
declarations = (
    family.location("age"),
    family.scale(),
)
model = SuperLSS(family, *declarations)

Passing declarations without * supplies one tuple where a predictor is expected. Declare each parameter exactly once. The family determines result order, so reordering the declarations does not reassign their meanings.

Move predictor controls onto the helper

family = GaussianLS()
model = SuperLSS(
    family,
    family.location("age", intercept=False, link="identity"),
    family.scale(),
)

Each helper accepts intercept and link. For interactions, declare the parent terms in the same predictor and include ti("left", "right") for a two-spline tensor, or interaction(existing_spec) for another supported interaction. ti supplies the interaction only; the parent smooths must be declared separately.

Keep parameter names in results and offsets

Tweedie’s construction helpers are mu, phi and p. Their names in prediction tables, offsets, penalty keys and saved models remain mean, dispersion and power. For example, a known addition to the mean’s log predictor belongs under offsets={"mean": values}.

Existing supported model artifacts still load through SuperLSS.from_bytes. The constructor migration does not change their parameter ordering or schema.

For a complete runnable example, follow Your first distributional model.