Evaluation Rubric
Overview

Evaluation Rubric

14 May 2026
1 min read

This is the rubric behind the partner track format. It is deliberately more concrete than “best AI demo wins” because partners need to see how teams reason, debug, and defend tradeoffs.

System Quality

Judges should be able to inspect the working path from input to output. For a retrieval system, that means query handling, chunking decisions, grounding, failure modes, and the evaluation set. For an edge system, it means latency, memory, battery assumptions, fallback behavior, and device constraints.

Constraint Handling

A strong submission makes constraints visible. Teams should state what they optimised for, what they refused to optimise for, and what would break if the problem were pushed harder.

Evidence

Useful evidence includes reproducible evaluations, error analysis, before/after comparisons, logs, screenshots, architecture diagrams, or live demos with clear test cases. The goal is not polish by itself. The goal is confidence that the team can reason about the system under pressure.

Communication

Finalists should explain the decision path without hiding behind jargon. A good presentation names the tradeoffs, shows the artifact, and gives judges enough detail to ask sharper questions.