i int64 1 30 | scenario stringlengths 69 219 | law_truth stringclasses 4
values | anchor stringlengths 26 79 | sov34 stringclasses 4
values | sov34_correct bool 2
classes | frontier_gemini stringclasses 2
values | frontier_correct bool 1
class | disagreement bool 2
classes | training_value stringclasses 2
values |
|---|---|---|---|---|---|---|---|---|---|
1 | A government platform scores citizens on trustworthiness using unrelated social behaviour, and the score restricts access to public services. | PROHIBITED | Art 5(1)(c) social scoring | HIGH_RISK | false | declined | false | true | medium |
2 | A wellness app assigns each employee a 'community contribution score' from their social media and off-duty behaviour; low scores reduce access to company benefits. | PROHIBITED | Art 5(1)(c) social scoring — benign framing, same practice | HIGH_RISK | false | declined | false | true | medium |
3 | A retailer scrapes CCTV and public webcam feeds to build a facial recognition database with no targeting criteria. | PROHIBITED | Art 5(1)(e) untargeted facial scraping | HIGH_RISK | false | declined | false | true | medium |
4 | Police run live facial recognition on a public square to identify anyone on a watchlist, with no judicial authorisation and no imminent threat. | PROHIBITED | Art 5(1)(h) real-time remote biometric ID in public | declined | false | declined | false | false | low |
5 | An AI ranks job applicants and filters CVs before any human sees them. | HIGH_RISK | Annex III 4(a) employment — recruitment | declined | false | declined | false | false | low |
6 | A hospital deploys a triage model that orders emergency patients by urgency. | HIGH_RISK | Annex III 5(a) essential services / emergency triage | LIMITED_RISK | false | declined | false | true | medium |
7 | A bank uses a model to decide consumer creditworthiness for mortgage applications. | HIGH_RISK | Annex III 5(b) creditworthiness | HIGH_RISK | true | declined | false | true | medium |
8 | A university admissions system scores applicants and determines who is admitted. | HIGH_RISK | Annex III 3(a) education access | declined | false | declined | false | false | low |
9 | A model assesses asylum applications and recommends grant or refusal to caseworkers. | HIGH_RISK | Annex III 7 migration and asylum | HIGH_RISK | true | declined | false | true | medium |
10 | An AI component controls pressure regulation in a municipal drinking-water network. | HIGH_RISK | Annex III 2 critical infrastructure safety component | declined | false | declined | false | false | low |
11 | An insurer uses AI for risk assessment and pricing of life and health insurance. | HIGH_RISK | Annex III 5(c) life/health insurance risk assessment | declined | false | declined | false | false | low |
12 | A retail website runs a customer-service chatbot that answers questions about orders. | LIMITED_RISK | Art 50(1) disclosure that the user interacts with an AI | LIMITED_RISK | true | declined | false | true | medium |
13 | A marketing tool generates photorealistic images of people for ad campaigns. | LIMITED_RISK | Art 50(2) synthetic content marking | HIGH_RISK | false | declined | false | true | medium |
14 | A newsroom tool produces synthetic voice-overs of presenters reading published articles. | LIMITED_RISK | Art 50(4) deep fake / synthetic audio disclosure | HIGH_RISK | false | declined | false | true | medium |
15 | A spam filter classifies inbound email for a company's staff mailboxes. | MINIMAL_RISK | No Annex III use case — employment context but not an employment decision | HIGH_RISK | false | declined | false | true | medium |
16 | A warehouse uses AI to optimise pallet placement and forklift routing for throughput. | MINIMAL_RISK | Industrial optimisation, no Annex III category, no safety component | declined | false | declined | false | false | low |
17 | A video game uses AI to adapt enemy difficulty to the player's skill. | MINIMAL_RISK | Recital ic — AI in video games, minimal risk | declined | false | declined | false | false | low |
18 | Software flags CVs containing keywords and forwards them to a recruiter, who reads every CV in full regardless and makes all decisions. It performs a narrow procedural task and does not materially influence the outcome. | MINIMAL_RISK | Art 6(3)(a) derogation — narrow procedural task, no material influence | declined | false | declined | false | false | low |
19 | A call-centre tool infers the emotional state of employees from their voice during calls and reports it to their managers. | PROHIBITED | Art 5(1)(f) emotion inference in the workplace | declined | false | declined | false | false | low |
20 | A security product categorises people from facial images to infer their ethnicity and political orientation. | PROHIBITED | Art 5(1)(g) biometric categorisation inferring protected attributes | HIGH_RISK | false | declined | false | true | medium |
21 | An AI is the safety component of a surgical robot, and that robot needs third-party conformity assessment under the Medical Devices Regulation. | HIGH_RISK | Art 6(1) — safety component of a regulated product; NOT via Annex III | declined | false | declined | false | false | low |
22 | A model is developed and used solely for scientific research and is never placed on the market or put into service. | MINIMAL_RISK | Art 2(6) — scientific research and development is out of scope | declined | false | declined | false | false | low |
23 | An AI profiles which residents of a city are likely to commit a crime, based purely on personality traits and background, with no link to objective verifiable facts. | PROHIBITED | Art 5(1)(d) predictive policing on profiling alone | declined | false | declined | false | false | low |
24 | A model that detects manufacturing defects on a bottling line is used to route faulty bottles to a reject bin. No person is assessed and no regulated product safety function depends on it. | MINIMAL_RISK | Industrial QC — no Annex III use, not an Art 6(1) safety component | declined | false | declined | false | false | low |
25 | A shopping app detects that a user's battery is about to die and raises prices, knowing the urgency makes them less able to compare alternatives. | PROHIBITED | Art 5(1)(a) exploitative/manipulative technique materially distorting behaviour | UNMEASURED | null | UNMEASURED | null | null | medium |
26 | A gambling platform identifies users showing signs of addiction and targets them with personalised 'one more bet' prompts at moments of detected distress. | PROHIBITED | Art 5(1)(b) exploitation of a vulnerability (addiction, emotional state) | UNMEASURED | null | UNMEASURED | null | null | medium |
27 | A warehouse deploys an AI system that continuously monitors pickers, allocates tasks, sets pace, and automatically issues disciplinary warnings when rates drop. | HIGH_RISK | Annex III 4(b) worker management — task allocation, monitoring, discipline | UNMEASURED | null | UNMEASURED | null | null | medium |
28 | A police force uses an AI system to assess the reliability of witness evidence during criminal investigations. | HIGH_RISK | Annex III 6(b) law-enforcement evaluation of evidence reliability | UNMEASURED | null | UNMEASURED | null | null | medium |
29 | A customer-service chatbot on a retail site answers in natural language with no statement anywhere that the user is talking to an AI system. | LIMITED_RISK | Art 50(1) transparency obligation — interaction with AI must be disclosed | UNMEASURED | null | UNMEASURED | null | null | medium |
30 | A factory uses an AI system to suggest optimal room-temperature setpoints for staff comfort; it touches no safety component and makes no decision about any person. | MINIMAL_RISK | No Annex III use case, no Art 6(1) safety component — comfort optimisation | UNMEASURED | null | UNMEASURED | null | null | medium |
COAI — EU AI Act tiering disagreement corpus
Start here — free verify: https://councilof.ai/gspc-verify/ (or the free MCP tool
verify_cardathttps://councilof.ai/mcp) · Full catalog: https://councilof.ai/catalog.json · Methodology: https://councilof.ai/methodology/ · Measurement, not certification.
A disagreement-mining corpus over EU AI Act tiering. Each row of sov_signal.jsonl
carries a scenario, the law_truth tier (PROHIBITED / HIGH_RISK / …), the anchor article that
settles it, what the local sov34 model answered and whether it was right, what a frontier model answered
and whether it was right, a disagreement flag and a training_value grade. The value is in the rows where
the two disagree — that is where the tiering is actually hard.
Superseded as a live bank; kept so inbound links resolve.
What is in this repository
| file | bytes | rows | what |
|---|---|---|---|
sov_signal.jsonl |
11822 | 30 | data (one JSON object per line) |
LICENSE |
11358 | file | |
METHODS_PAPER.md |
7792 | card | |
spine_accuracy.json |
6997 | JSON document | |
LEADERBOARD.md |
6227 | card | |
final_e2e.json |
2366 | JSON document | |
all_models_enhanced.json |
2332 | JSON document | |
sov_family_board.json |
2219 | JSON document | |
README.md |
1517 | card | |
full_15dim.json |
1120 | JSON document | |
sov-sovereign-v4_latest.json |
975 | JSON document | |
manifest.jsonl |
33 | derived at 2026-09-04T04:33:29Z |
manifest.jsonl is derived from this repository's own file tree at 2026-09-04T04:33:29Z; it lists every file with its
size, its blob hash and a direct URL, so an agent can enumerate the repo without cloning it.
Citation
@misc{csoai_gspc_coai_bench,
title = {COAI — EU AI Act tiering disagreement corpus},
author = {{CSOAI Ltd}},
year = {2026},
note = {DOI record unavailable since 29 Sep 2026 (Zenodo account blocked; appeal pending)},
howpublished = {Hugging Face Hub, \url{https://huggingface.co/datasets/csoai/coai-bench}},
note = {Live board: https://councilof.ai/api/gspc. Measurement, not certification.}
}
Objections, contact and corrections
To object to a row, ask for a re-check, request a correction or ask for a record to be withdrawn, email nicholas@csoai.org or use the appeals and dispute route at https://councilof.ai/dispute/. Corrections are listed in the corrections ledger at https://councilof.ai/api/corrections, with what changed and when. CSOAI Ltd (company no. 16939677, England and Wales) is the accountable publisher.
How to cite
CSOAI Ltd (Council of AI). COAI — EU AI Act tiering disagreement corpus. 2026. Hugging Face dataset csoai/coai-bench. https://huggingface.co/datasets/csoai/coai-bench
@misc{csoai_coai_bench,
title = {COAI — EU AI Act tiering disagreement corpus},
author = {{CSOAI Ltd}},
year = {2026},
howpublished = {Hugging Face dataset, https://huggingface.co/datasets/csoai/coai-bench},
note = {Corrections: https://councilof.ai/api/corrections}
}
Licence: CC-BY-4.0. Attribute Council of AI, CSOAI Ltd (16939677), https://councilof.ai.
Corrections and verification
- Corrections ledger (signed): https://councilof.ai/api/corrections. Corrections to CSOAI's published records are logged there with what changed and when.
- Verify a signed record yourself, free and without an account: https://councilof.ai/gspc-verify/ (step by step: https://councilof.ai/signed/HOW-TO-VERIFY.md).
- Conformance kit for signed-receipts/v1, with test vectors for implementers: https://councilof.ai/spec/signed-receipts/v1/conformance/
Reuse and provenance
Download one published file at the recorded revision:
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id='csoai/coai-bench', repo_type='dataset',
filename='sov_signal.jsonl',
revision='6f7b8cbdf420b81bf8321dbf1ce5bd90af10c7ef',
)
print(path)
Use the file's observation date, scope and licence when reusing it. Repository updates and downloads are not new measurements.
- Current GSPC board and methodology.
- Verify supported signed records and read corrections.
- Earlier page notes and dated integration announcements.
Measurement is not certification. A repository, signed record or passing software test does not by itself establish current board admission, product acceptance or benchmark leadership.
- Downloads last month
- 845