An Open Pipeline and Dashboard for Systemic-Risk Evidence under the EU AI Act's Code of Practice
Jacob T. Emmerson, Phuong-Anh Nguyen-Le, Ronan Romano, Wilber Sean V. Anterola, Yann Billeter, Zhijing Jin
cs.AI
Quick verdict
Open pipeline scoring 18 models on four EU AI Act systemic-risk categories; transparent and reproducible, but small validation samples undercut some claims.
Summary
The paper presents the EuroSafeAI Systemic Risk Index, an open evaluation pipeline and dashboard that maps 19 public safety benchmarks (1,188 items) onto four systemic-risk categories from the EU GPAI Code of Practice: CBRN, cyber offense, harmful manipulation, and loss of control. Built on the UK AISI Inspect framework, it evaluates 18 models under a baseline (excluded from scoring), four harm-preserving perturbations, and generated agentic scenarios, graded by reference answers, deterministic rules, or two LLM judges. The dashboard exposes provider- and model-level views, average vs. worst-case aggregation, a capability-weight slider, and per-rating benchmark provenance. Worst-case aggregation lowers scores by 14.76 to 36.59 points across the 18 models. LLM-judge agreement with human raters (κ=0.78–0.82, n=95) approaches human–human agreement (κ=0.82), and a blind audit finds 83% of sampled transformations preserve the original harm, though harmful-manipulation items fall to 0.50 under paraphrase/register rewrites. A 21-participant convenience-sampled survey reports strong perceived score intelligibility (20/21 on score meaning and direction) but weak understanding of evidential limits (10/21). Main limitations are small validation samples (some κ cells based on 4–5 items, 8 pairs per audit cell), judgment-based benchmark-to-risk mapping, baseline exclusion that may favor contaminated models, and self-report-only user evidence. Code and data are public.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An open pipeline plus dashboard can organize 19 public benchmarks (1,188 items) into the four EU GPAI Code of Practice systemic-risk categories and make every aggregate rating traceable to benchmark evidence.
- Worst-case aggregation lowers scores by 14.76–36.59 points across 18 models relative to mean aggregation, exposing weak individual responses hidden by averages (Table 2).
- LLM judges agree with blind human raters at levels comparable to human–human agreement (κ=0.78–0.82 vs 0.82 on 95 sampled responses), supporting automated rubric grading (Table 3).
- A blind audit finds 83% of sampled transformations preserve the original harm, with harmful-manipulation items weakest (0.50 under paraphrase/register rewrites, Table 4).
- In a 21-participant survey, users readily understand score semantics (20/21) but poorly understand what scores do not represent (10/21, Tables 1/5).
Strengths
Full evidence provenance by design
§2 and Figures 1/5: every cell expands into the contributing benchmarks with descriptions and source links; aggregation rule (mean vs worst-case) and capability weighting are exposed as user controls rather than hidden defaults — directly addresses the black-box leaderboard critique in §1.
Genuine reproducibility commitment
Code, data, and results are public (footnote 1); pipeline built on open Inspect framework (§3); all transformation prompts published (Appendix B); Appendix C gives a fully worked end-to-end example including model outputs, so the scoring semantics are inspectable.
Both automated components are validated against humans
Honest boundary-setting
Quantified aggregation sensitivity across 18 models
Weaknesses
Validation samples too small for per-cluster claims
Baseline exclusion introduces an unquantified bias
Worst-case effect conflates behavior with judge noise
Subjective mapping and sampling pipeline
User evidence is perception-only and small
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper introduces the EuroSafeAI Systemic Risk Index: an open evaluation pipeline (built on the UK AISI Inspect framework) plus a static React dashboard that reorganizes 1,188 items from 19 public safety benchmarks into the four systemic-risk categories of the EU GPAI Code of Practice (CBRN, cyber offense, harmful manipulation, loss of control). The mapping runs benchmarks → harms → risk categories via the MIT Risk Repository, curated by author judgment (§3.1). Three central claims: (i) aggregate ratings can be made fully traceable to benchmark evidence; (ii) aggregation choices matter — worst-case aggregation drops scores by 14.76–36.59 points across 18 models (Table 2); (iii) automated grading is defensible — LLM judges agree with humans at κ=0.78–0.82, near human–human agreement of 0.82 (Table 3), and 83% of transformations preserve harm (Table 4). A 21-person survey (Tables 1/5) supports perceived intelligibility. This is a system/infrastructure paper, not a claim about any specific model's safety, and it explicitly disclaims legal-compliance use (Limitations, Ethical Considerations).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Systemic Risk Index?
It is an open evaluation pipeline and interactive dashboard that scores AI models on four systemic-risk categories from the EU GPAI Code of Practice — CBRN, cyber offense, harmful manipulation, and loss of control — using 19 public benchmarks (1,188 items) and lets any user trace each score back to its benchmark evidence.
What did the paper find about the models it evaluated?
Across 18 models, average-aggregated safety scores range roughly from 45 to 90. Switching to worst-case aggregation drops scores by 14.76 to 36.59 points, showing that weak responses on individual transformed items are hidden by averages; open and smaller models (e.g., Mistral, DeepSeek, Llama variants) show the largest drops.
Can the results be reproduced?
The authors state that code, data, and results are public on GitHub, the pipeline runs on the open Inspect framework, and all perturbation and scenario prompts are published in the paper's appendix, so evaluations can be rerun — though live model APIs and judge models can drift over time.
How reliable are the automated (LLM) judges?
Two LLM judges agreed with blind human raters at Cohen's κ of 0.78–0.82 on 95 sampled responses, close to the human–human agreement of κ=0.82. However, per-risk-category human samples were small (as few as 4–5 items), and one judge refused to score 12 sampled responses.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–6 (dashboard, pipeline diagram, expanded leaderboard) are images and were assessed only via captions and the questionnaire text, not visual inspection.
- The names of the 19 constituent benchmarks are not enumerated in the extracted text (only example sources like SocialHarmBench appear), so coverage-balance judgments are limited.
- The demo URL in footnote 1 appears possibly mangled by extraction (https://safe.eu/index); code availability is taken from the stated GitHub repository link.
- Model names/versions in Table 2 (e.g., GPT-5.6-Luna, Gemini 3.6 Flash) could not be verified against the text's citations.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.