Preserve model vs population semantics in performance tables

[uriahf]

Goal

Align Python performance-table presentation with the merged evaluation semantic contract without changing prepared-performance-data schemas or public APIs.

Problem

Both Python table renderers currently rename the generic compatibility field reference_group to Model. For keyed population/evaluation inputs, those groups are populations, so the rendered table can label Population A under a Model heading. R preserves the known semantic role and renders Model or Population accordingly.

Change

High-level create_performance_table() and create_performance_table_times() now derive the known evaluation role from the existing internal evaluation metadata and attach it only as transient render-path metadata.

• shared outcomes -> groups are models in one shared population -> render Model;
• keyed outcomes/times -> groups are distinct populations/evaluations with unknown model identity -> render Population;
• bare render_performance_table(prepared_data) retains the existing default behavior because the original input identity is no longer available there.

Both Great Tables and Reactable use the transient role only for rendering/group-label semantics and coloring.

Compatibility

• no changes to prepare_performance_data() or prepare_performance_data_times() return schemas;
• no new public arguments;
• no statistical changes;
• no changes to row identity, ordering semantics, or calculations;
• no R or rtichoke_viz changes.

Tests

Adds regression coverage that shared-outcome create paths propagate model role, keyed population paths propagate population role, and prepare_performance_data_times() itself remains free of the transient render metadata.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-21 14:48 UTC