Concepts

Warning

Pre-implementation. This section describes proposed contracts. Behavior is subject to change before code lands.

Who should read this. Anyone who needs the mental model behind EMMA’s scores.

Before you start. Read the home page first; this section assumes the SUPERB-for-RF framing.

What this section is not. It is not a contract reference; see Reference and API Reference for exact schemas.

Orientation

EMMA (Electromagnetic Multi-task Model Assessment) rests on four design pillars: (1) generalization is the score, not in-distribution accuracy; (2) the canonical input is raw, phase-coherent, multi-antenna I/Q (in-phase and quadrature), which localization and beam tasks physically require; (3) the benchmark is reproducible by construction, with synthetic scenes validated against real captures; and (4) governance is neutral by design, ring-fenced from any one model team.

The pages below expand each pillar from a user’s point of view. Start with Signal as a modality for the structural template and what EMMA changes. Tasks and Generalization define what is scored and how. Data model, Sim-to-real gap, and Reproducibility cover how data is shaped and pinned. Neutrality and Leaderboard close the loop on governance and reporting.

The end-to-end evaluation pipeline and the dataset-generation internals are in the developer documentation.