Evaluation Cards: An Interpretive Layer for AI Evaluation Reporting

2026-06-08

AI Summary

This work addresses the lack of consistency in current AI evaluation reporting, which impedes cross-source comparison, evidence traceability, and identification of missing information. To resolve this, we propose EvalCards—the first composable reporting framework spanning the full evaluation lifecycle—dynamically adapting to diverse audience needs for interpretability by integrating benchmark metadata, evaluation run data, and model metadata. Grounded in an analysis of 52 papers and interviews with 10 stakeholders, we define four classes of explanatory signals: reproducibility, documentation completeness, provenance risk, and score comparability, and develop a corresponding large-scale extraction and monitoring infrastructure. The system currently encompasses 5,816 models, 635 benchmarks, and 101,843 evaluation results, uncovering systemic deficiencies in contemporary AI evaluation practices.

Abstract

AI evaluation results are produced at scale but reported inconsistently across leaderboards, model cards, benchmark papers, and company blogs. The cost is interpretive: readers cannot reliably compare results across sources, identify what a report omits, or trace an aggregate claim to its underlying evidence. Recent efforts address isolated components but leave three gaps: they cover only narrow slices of the evaluation lifecycle and do not compose into a single interpretable record; they specify static representations that do not differentiate the questions different stakeholders bring to the same evidence; and they remain proposals on paper, lacking the extraction infrastructure required for adoption at scale. We present \EvalCards{}, an operational reporting layer that composes benchmark metadata, evaluation run data, and model metadata into a unified record. We (1) derive a reporting schema from a structured review of 52 papers and 10 stakeholder interviews, (2) implement four interpretive signals (reproducibility, documentation completeness, provenance and risk, and score comparability), rendered through reader modes calibrated to research and non-research audiences, and (3) deploy a monitoring tool that applies \EvalCards{} across 5,816 models, 635 benchmarks, and 101,843 results, surfacing systematic gaps in current reporting practice.

Problem

• AI evaluation
• reporting inconsistency
• interpretability
• evaluation lifecycle
• evidence traceability

Innovation

• Evaluation Cards
• Interpretability
• AI Evaluation
• Metadata Integration
• Stakeholder-Centric Reporting