Skip to content
Store / AI Governance / AI System Accuracy Checklist
AI Governance AI Act Arts. 13 & 15

AI System Accuracy Checklist (AI Act Arts. 13 & 15)

The 13 mandatory elements, with cross-references to the risk management, post-market monitoring and incident reporting work you probably already have — plus a starter template for the 3 that almost nobody has covered yet.

Based on AESIA Guide 9 (Spain's AI Supervision Agency), born out of its regulatory sandbox and adapted to the Digital Omnibus — the practical reference standard for compliance in Spain.
€29
VAT included · one-time payment
Coming soon → ← Back to store
Referenced article by article with the AI Act, GDPR and DAMA-DMBOK

What's included

  • Excel file with the 9 accuracy evaluation areas, each with a guiding question and a practical recommendation
  • Accuracy log: model, version, declared metric, evaluation set and next re-evaluation date
  • Subgroup breakdown: the table that reveals whether your overall accuracy hides worse performance for a particular group
  • Self-assessment by area, with owner and last review date
  • PDF guide · Legal Notice

Why a bare percentage doesn't satisfy the requirement

Almost anyone can quote an accuracy number. The AI Act asks for four more things: justifying why that metric and not another, demonstrating the model didn't memorize its training set, checking accuracy is comparable across groups of people, and declaring it in the instructions for use (Art. 13). A percentage on a slide satisfies none of the four.

The 9 areas it covers

  • Defining accuracy for your case — whether a false positive costs the same as a false negative, or not.
  • Preprocessing — transformations that are documented and reproducible in production too.
  • Overfitting — strict separation between training, validation and test sets.
  • Metric selection — which ones, and why those.
  • Objective function — what the model actually optimizes for, and whether it matches what matters.
  • Subgroup breakdown — where it fails more, and for whom.
  • Declaration (Art. 13) — that it appears in the instructions for use, not just an internal report.
  • Assurance and traceability — that the evaluation can be reproduced: data, version, seed, environment.
  • Periodic re-evaluation — how often it's re-measured and what threshold forces action.

Fits what you already have

If you use the Model Card Template, this checklist is what fills its metrics section with substance: the Model Card tells you what to write down, this document tells you how to arrive at that number and how to defend it. And it complements the Robustness Checklist, which covers the other half of Art. 15.