Layer 7: Interpretation / Importance

Layer 7 explains model forecasts through importance, attribution, marginal effects, lineage aggregation, and transformation attribution. It is default off and uses graph-form YAML.

Contract

Inputs:

  • l4_model_artifacts_v1;

  • l4_forecasts_v1;

  • l3_features_v1;

  • l3_metadata_v1;

  • l5_evaluation_v1;

  • optional l6_tests_v1;

  • optional L1 data/regime metadata.

Outputs:

  • l7_importance_v1;

  • l7_transformation_attribution_v1 when transformation attribution is used.

Sub-Layers

Slot

Purpose

L7.A

importance DAG body

L7.B

output shape and export axes

Compatibility Rules

  • Tree SHAP and tree-native importance require tree model families.

  • Linear SHAP, coefficient importance, and forecast decomposition require linear model families.

  • Deep attribution ops require neural-network model families.

  • VAR-specific ops require VAR or BVAR families.

  • mrf_gtvp requires macroeconomic_random_forest.

  • MCS-filtered sources require active L6 MCS.

  • L7 output axes are not sweepable.

Example

7_interpretation:
  enabled: true
  nodes:
    - {id: src_model, type: source, selector: {layer_ref: l4, sink_name: l4_model_artifacts_v1, subset: {model_id: xgb_full}}}
    - {id: src_X, type: source, selector: {layer_ref: l3, sink_name: l3_features_v1, subset: {component: X_final}}}
    - {id: src_l3_meta, type: source, selector: {layer_ref: l3, sink_name: l3_metadata_v1}}
    - {id: shap, type: step, op: shap_tree, params: {model_family: xgboost}, inputs: [src_model, src_X]}
    - {id: lineage, type: step, op: lineage_attribution, params: {level: pipeline_name}, inputs: [shap, src_l3_meta]}
  sinks:
    l7_importance_v1: {global: shap, lineage: lineage}
  fixed_axes:
    figure_type: auto

See encyclopedia

For the full per-axis × per-option catalogue (every value with its OptionDoc summary, when-to-use / when-NOT, references), see encyclopedia/l7/.