Layers¶

Built-in layers:

IdealPointLayer
PublicOpinionLayer
LobbyingLayer
MediaPressureLayer
PartyDisciplineLayer
GovernmentAgendaLayer

Optional:

neural layer support (include_neural + custom factory).

Custom composition¶

Use layer_names and layer_overrides to build an explicit layer pipeline.

What each layer usually represents¶

Layer	Typical mechanism	Common tuning focus
`IdealPointLayer`	actor-bill preference proximity	baseline preference shape
`PublicOpinionLayer`	electorate-level pressure	`public_support`
`LobbyingLayer`	targeted influence pressure	lobbyist count/strength
`MediaPressureLayer`	salience and narrative pressure	media intensity
`PartyDisciplineLayer`	party-line conformity	discipline strength
`GovernmentAgendaLayer`	agenda access/priority effects	executive/speaker alignment

Minimal layer-first workflow¶

Start with ideal point only.
Add one extra layer.
Keep seed and aggregation fixed.
Compare pass_rate deltas.
Repeat for next layer.

This yields cleaner attribution than enabling many layers at once.

Explicit order and overrides example¶

from policyflux import IntegrationConfig, LayerConfig, build_engine

config = IntegrationConfig(
    num_actors=100,
    policy_dim=2,
    iterations=200,
    seed=7,
    layer_config=LayerConfig(
        include_ideal_point=True,
        include_public_opinion=True,
        include_party_discipline=True,
        public_support=0.6,
        party_discipline_strength=0.5,
        layer_names=[
            "ideal_point",
            "public_opinion",
            "party_discipline",
        ],
        layer_overrides={
            "public_opinion": {"support": 0.62},
        },
    ),
)

engine = build_engine(config)
engine.run()
print(engine.pass_rate)

Neural layer notes¶

When using neural layers:

install optional dependencies first,
keep a deterministic baseline for comparison,
document model artifacts and runtime assumptions for reproducibility.

Troubleshooting¶

Hard-to-interpret output changes

Reduce active layers and reintroduce them one by one.

Sensitivity to layer order

Use explicit layer_names and version-control this order across experiments.