Template
EU AI Act Article 15 - Accuracy and Robustness Template
Article 15 is where performance claims stop being slogans and start needing evidence. The documentation has to show quality signals, robustness and security constats, version-to-version comparison, and a clear point where the system no longer meets the provider's accepted threshold.
Who supplies what
What Article 15 expects, what the toolkit can add, and what your team still owns
Without connecting your agent, the toolkit can only expose the quality and security outputs already present in the generated package. After adapter integration, it can keep those outputs tied to real runs, version changes, and repeatable evidence over time.
| Article 15 expects | Toolkit with generated package only | Toolkit after adapter integration | Still written and approved by your team |
|---|---|---|---|
| Quality and accuracy signals | Shows pass rate, execution-quality status, and case outcomes already present in the rapport de comparaison. | Keeps those signals tied to the connected agent's real runs, versions, and case-level evidence. | Define which task metrics matter, what score is acceptable, and what counts as success for your use case. |
| Robustness and security constats | Shows current security constats and risk levels already present in the package. | Can keep those constats linked to the actual runs, tool behavior, and case outputs that triggered them. | Explain which robustness or security constats are acceptable, and which require remediation before the system is relied on. |
| Version-to-version comparison | Shows baseline versus new comparison already built into the rapport de comparaison. | Lets that comparison stay attached to real run inputs, outputs, and retained artifacts over time. | Decide whether the chosen cases are sufficient to justify the claim you want to make. |
| A documented performance threshold | Shows current execution-quality status and case outcomes already in the package. | Can keep the threshold evidence tied to recurring runs and monitoring history as the system changes. | Set the final threshold for acceptance, escalation, or block and approve that decision. |
Manual fields
What your team still defines around Article 15
The toolkit can generate the evidence dossier. It does not decide what quality level is acceptable for your domain or whether the chosen case set is sufficient.
| What you define | When you define it | Practical format to use |
|---|---|---|
| Task-specific success threshold | Before you rely on the package to support performance or conformity claims. | A short threshold statement or table naming the metric, target, and minimum acceptable result. |
| Why this case suite is sufficient | When you claim the current evidence is enough to support a quality statement. | A short rationale naming which cases cover the main risks, failure modes, or edge conditions. |
| Business impact of remaining failures | When open regressions, degraded execution, or risky cases remain in the package. | One or two sentences describing what those failures mean for your deployment and users. |
| Accepted robustness and security risk | Before sign-off if any significant issue remains open. | A short acceptance, block, or escalate decision plus the owner who approved it. |
| Broader performance claims | Whenever you want to claim more than the package directly proves. | A plain-language note separating what the evidence proves from any wider claim your organization wants to make. |
How this fits into the full package
Article 15 is documented through comparison, integrity, and performance outputs
The files below are the main evidence surfaces for Article 15. Their layout is the toolkit's structured format for these requirements, not an EU-mandated form.
| File in the package | Why it matters for Article 15 | Open file |
|---|---|---|
| Compare report - compare-report.json | Main machine-readable comparison showing pass rates, execution quality, security constats, and highlighted cases. | Open file |
| EU AI Act evidence report - report.html | Human-readable evidence report used to inspect the same quality and regression outputs. | Open file |
| Manifest - artifacts/manifest.json | Integrity index anchoring the files that support the Article 15 claims. | Open file |
| Article 9 risk register - article-9-risk-register.json | Shows how serious performance or security issues can reopen or update risk entries. | Open file |
FAQ
Frequently asked questions
What is the Article 15 - Accuracy and Robustness template page for?
This page explains what the Article 15 - Accuracy and Robustness template section usually needs to cover inside an EU AI Act package, which parts can be supported by structured evidence, and which parts remain owned by the provider, deployer, or legal reviewer.
Does this Article 15 - Accuracy and Robustness template page create a complete EU AI Act package?
No. It is one section-level guide. A review-ready package still needs the selected role and scope, the other applicable article-level sections, linked technical evidence, owner completion, and final human review.
What evidence should be linked to the Article 15 - Accuracy and Robustness template section?
Use evidence that is current, traceable to the relevant system version, and connected to real runs, retained records, or approved procedures. A weak package relies only on narrative text; a stronger package links the narrative to reproducible artifacts.
Who should review the Article 15 - Accuracy and Robustness template section before handoff?
The technical owner should confirm the system facts and evidence links, while compliance or legal reviewers confirm whether the section is sufficient for the selected EU AI Act role, scope, and conformity path.