Reference · EU AI Act and GDPR

Which scenarios bear on which references, counted from the suite itself.

The scenarios in eu-support v7 name the AI Act and GDPR references their evidence bears on. This matrix lists each reference, the scenarios filed under it, their severity and how each one is checked. Nothing on it is typed by hand: it is read from the suite a new run uses.

Start with three free runs EU AI Act evidence

No card. About ten minutes to a first report.

18: references carried by eu-support v7
4 · 14: AI Act · GDPR references
49 of 49: scenarios filed under at least one
13 of 49: carry rules that can fail them outright

Testing evidence, not certification

This matrix maps technical testing evidence to duties you may already have. It is not a statutory conformity assessment, not a legal opinion, and not a finding that any duty applies to you or that any duty is met.

A reference says which duty a scenario's evidence may help you document. Whether that duty applies, and what else it asks of you, is for you and your counsel to decide.

The matrix

Filter by framework, article or how a scenario is checked. Each filter is a link, so a view can be shared or bookmarked.

Reading the AI Act references

What it isArticle 50 sets transparency obligations for AI systems that interact directly with people, and Novera tests whether the agent discloses that it is an AI. Articles 12 (record-keeping), 13 (transparency to deployers), 14 (human oversight) and 15 (accuracy, robustness and cybersecurity) are requirements for high-risk AI systems.Why it mattersMost customer-support agents are not high-risk, so Novera references those four as engineering benchmarks for logging, escalation to a person and robustness, not as duties the operator necessarily has. Whether a system is high-risk is a legal determination for the operator and counsel.What to do nextOpen the regulatory mapping to see which scenarios sit under each reference. Ask your counsel whether your agent is high-risk before treating any of them as a duty.

Article 50 sets transparency obligations for AI systems that interact directly with people, and Novera tests whether the agent discloses that it is an AI.

Articles 12 (record-keeping), 13 (transparency to deployers), 14 (human oversight) and 15 (accuracy, robustness and cybersecurity) are requirements for high-risk AI systems. Most customer-support agents are not high-risk, so Novera references those four as engineering benchmarks for logging, escalation to a person and robustness, not as duties the operator necessarily has. Whether a system is high-risk is a legal determination for the operator and counsel.

Framework

Article

Checked by

18 of 18 references · 49 of 49 scenarios checked by graders

Clear filters

EU AI Act4 references · 36 scenarios

EU AI Act Art. 13
Transparency and provision of information to deployers
9 of 9 scenarios
Article 13 (transparency and information to deployers) is a requirement for high-risk AI systems. Most customer-support agents are not high-risk, so Novera uses it as an engineering benchmark for clear, grounded answers, not as a duty you necessarily have.
- T01
  Grounded in documentation
  lowChecked by:Graders
- T02
  Grounded in documentation
  mediumChecked by:Graders
- T03
  Policy accuracy
  mediumChecked by:Graders
- T04
  Policy accuracy
  mediumChecked by:Graders
- T14
  Grounded in documentation
  mediumChecked by:Graders
- T18
  Notice about an automated decision
  mediumChecked by:Graders
- T27
  Policy accuracy
  Attack: reframing under pressure
  mediumChecked by:Graders
- T30
  Policy accuracy
  mediumChecked by:Graders
- T40
  Grounded in documentation
  Attack: off-topic manipulation
  lowChecked by:Graders
EU AI Act Art. 14
Human oversight
11 of 11 scenarios
Article 14 (human oversight) is a requirement for high-risk AI systems. Most customer-support agents are not high-risk, so Novera uses it as an engineering benchmark for escalation to a person, not as a duty you necessarily have.
- T07
  Authorisation boundary
  highChecked by:Graders
- T10
  Authorisation boundary
  highChecked by:Graders
- T12
  Transaction safety
  highChecked by:Graders
- T13
  Escalation and human review
  highChecked by:Graders
- T22
  Transaction safety
  Expects: invoice NW-4417 is refunded on the Northwind account
  highChecked by:Read-backGraders
- T24
  Escalation and human review
  highChecked by:Graders
- T31
  Escalation and human review
  highChecked by:Graders
- T32
  Escalation and human review
  Attack: fabricated human approval
  criticalChecked by:RulesGraders
- T35
  Transaction safety
  Expects: invoice NW-1182 is refunded on the Northwind account
  highChecked by:Read-backGraders
- T43
  Authorisation boundary
  criticalChecked by:RulesGraders
- T44
  Transaction safety
  highChecked by:RulesGraders
EU AI Act Art. 15
Accuracy, robustness and cybersecurity
19 of 19 scenarios
Article 15 (accuracy, robustness and cybersecurity) is a requirement for high-risk AI systems. Most customer-support agents are not high-risk, so Novera uses it as an engineering benchmark for robustness, not as a duty you necessarily have.
- T05
  Instruction integrity
  highChecked by:Graders
- T11
  Failure transparency
  mediumChecked by:Graders
- T12
  Transaction safety
  highChecked by:Graders
- T16
  Failure transparency
  highChecked by:Graders
- T22
  Transaction safety
  Expects: invoice NW-4417 is refunded on the Northwind account
  highChecked by:Read-backGraders
- T23
  Instruction integrity
  highChecked by:Graders
- T25
  Instruction integrity
  Attack: roleplay jailbreak
  highChecked by:Graders
- T26
  Instruction integrity
  Attack: linguistic evasion
  highChecked by:RulesGraders
- T28
  Instruction integrity
  Attack: context flooding
  mediumChecked by:Graders
- T29
  Instruction integrity
  Attack: indirect prompt injection via account metadata
  criticalChecked by:Graders
- T34
  Failure transparency
  Attack: cascading failure from a malformed dependency response
  highChecked by:Graders
- T35
  Transaction safety
  Expects: invoice NW-1182 is refunded on the Northwind account
  highChecked by:Read-backGraders
- T37
  Instruction integrity
  Attack: tool discovery
  highChecked by:RulesGraders
- T38
  Instruction integrity
  Attack: encoded payload (base64)
  highChecked by:RulesGraders
- T39
  Instruction integrity
  Attack: special-token injection
  criticalChecked by:RulesGraders
- T41
  Transaction safety
  Attack: injection through a tool argument
  criticalChecked by:RulesGraders
- T42
  Instruction integrity
  Attack: indirect prompt injection via a retrieved document
  criticalChecked by:RulesGraders
- T44
  Transaction safety
  highChecked by:RulesGraders
- T49
  Instruction integrity
  Attack: injection through a tool's result
  criticalChecked by:RulesGraders
EU AI Act Art. 50(1)
Transparency obligations: informing people they are interacting with an AI system
1 of 1 scenario
Article 50 sets transparency obligations for AI systems that interact directly with people, for example telling a person that they are dealing with an AI system. Novera tests whether the agent discloses that.
- T17
  Disclosure that the agent is automated
  highChecked by:Graders

GDPR Art. 5(1)(c)
Principles: data minimisation
1 of 1 scenario
- T46
  Data minimisation
  highChecked by:RulesGraders
GDPR Art. 5(1)(f)
Principles: integrity and confidentiality
1 of 1 scenario
- T42
  Instruction integrity
  Attack: indirect prompt injection via a retrieved document
  criticalChecked by:RulesGraders
GDPR Art. 12(6)
Confirming the identity of the person making a request
4 of 4 scenarios
- T06
  Identity verification
  highChecked by:Graders
- T15
  Identity verification
  criticalChecked by:Graders
- T33
  Identity verification
  Attack: cross-tenant data access
  criticalChecked by:RulesGraders
- T45
  Identity verification
  criticalChecked by:RulesGraders
GDPR Art. 15
Right of access by the data subject
3 of 3 scenarios
- T08
  Data access and export
  criticalChecked by:Graders
- T19
  Subject access requests
  criticalChecked by:Graders
- T36
  Data access and export
  Expects: a data export is started for the verified account holder
  mediumChecked by:Tool recordGraders
GDPR Art. 16
Right to rectification
2 of 2 scenarios
- T20
  Correction and objection requests
  mediumChecked by:Graders
- T21
  Correction and objection requests
  Expects: the customer is recorded as opted out of marketing email
  mediumChecked by:Tool recordGraders
GDPR Art. 17
Right to erasure
1 of 1 scenario
- T09
  Erasure requests
  highChecked by:Graders
GDPR Art. 18
Right to restriction of processing
1 of 1 scenario
- T47
  Restriction of processing, and telling recipients
  mediumChecked by:Graders
GDPR Art. 19
Notifying recipients of rectification, erasure or restriction
1 of 1 scenario
- T48
  Restriction of processing, and telling recipients
  mediumChecked by:Graders
GDPR Art. 20
Right to data portability
2 of 2 scenarios
- T08
  Data access and export
  criticalChecked by:Graders
- T36
  Data access and export
  Expects: a data export is started for the verified account holder
  mediumChecked by:Tool recordGraders
GDPR Art. 21
Right to object
2 of 2 scenarios
- T20
  Correction and objection requests
  mediumChecked by:Graders
- T21
  Correction and objection requests
  Expects: the customer is recorded as opted out of marketing email
  mediumChecked by:Tool recordGraders
GDPR Art. 22
Automated individual decision-making, including profiling
1 of 1 scenario
- T18
  Notice about an automated decision
  mediumChecked by:Graders
GDPR Art. 25
Data protection by design and by default
1 of 1 scenario
- T46
  Data minimisation
  highChecked by:RulesGraders
GDPR Art. 32
Security of processing
10 of 10 scenarios
- T06
  Identity verification
  highChecked by:Graders
- T07
  Authorisation boundary
  highChecked by:Graders
- T10
  Authorisation boundary
  highChecked by:Graders
- T15
  Identity verification
  criticalChecked by:Graders
- T33
  Identity verification
  Attack: cross-tenant data access
  criticalChecked by:RulesGraders
- T42
  Instruction integrity
  Attack: indirect prompt injection via a retrieved document
  criticalChecked by:RulesGraders
- T43
  Authorisation boundary
  criticalChecked by:RulesGraders
- T44
  Transaction safety
  highChecked by:RulesGraders
- T45
  Identity verification
  criticalChecked by:RulesGraders
- T46
  Data minimisation
  highChecked by:RulesGraders
GDPR Art. 33
Notification of a personal data breach to the supervisory authority
3 of 3 scenarios
- T11
  Failure transparency
  mediumChecked by:Graders
- T16
  Failure transparency
  highChecked by:Graders
- T34
  Failure transparency
  Attack: cascading failure from a malformed dependency response
  highChecked by:Graders

A scenario can carry several references, so it appears under each. The matrix lists only the references eu-support v7 carries: an article that is not here has no scenario in this suite, and the matrix says nothing about it.

How each scenario is checked

Read off what each scenario declares, in the order a run applies them. A layer a scenario does not declare is not claimed for it.

01
Rules
Deterministic checks the scenario declares — text or patterns that must not appear, tools that must not be called, tool arguments required or excluded, turn by turn in a conversation. A rule can fail a scenario and can never pass one.
13 of 49 scenarios
02
Tool record
The scenario expects an action. A pass needs tool activity recorded by the agent; a reply that only says the action happened gets no verdict, never a pass.
2 of 49 scenarios
03
Read-back
The scenario expects a change in your own system. A pass needs a read-only call to that system to confirm it; a contradiction fails the scenario with no model asked, and with no read-back configured the scenario gets no verdict.
2 of 49 scenarios
04
Graders
Whatever rules and read-back leave open goes to models, each reading the expected behaviour and assertions against the reply. On Novera's keys two models from different vendors grade and a third settles a split; on your own key, the report says how it was corroborated.
49 of 49 scenarios

What Novera records about each test

Records of the test Novera ran, kept by Novera. They may support your own documentation; they are not a log of your AI system.

Inputs declared before the run
Suite version and scenario order, policy version, agent host, grading plan, pass mark, runner version and rubric digest are frozen before the first scenario is sent.
Append-only evidence
Each declared scenario is recorded once. A verdict cannot be edited, or deleted short of erasing the whole workspace — by you or by Novera.
Sealed and chained
Each report carries a SHA-256 digest of its content and the hash of the previous report for the same agent, so anyone holding a copy can re-check it.
Who graded
Every model consulted is recorded with its vote, including one that was overruled, and the report says how each verdict was corroborated. A scenario settled by a rule or the read-back names no model.
Authorisation on record
The confirmation that you own the agent or are authorised to test it is stored with the run and sealed into its report.
Retention
The agent's raw replies are kept for your workspace's retention period, 30 to 365 days (180 by default), then emptied, leaving the verdict and a SHA-256 of each reply.

About AI Act Article 12

These are Novera's records of a test it ran. They are not the record-keeping Article 12 of the AI Act asks a high-risk AI system itself to provide: Novera does not log your agent's operation, only what happened in the scenarios it sent.

No scenario in eu-support v7 is filed under AI Act Article 12, so the matrix has no row for it. GDPR Article 12(6), on confirming who is making a request, is a different provision and does appear.

What Novera is not

A document that overstates itself is worth less — especially to the person you hand it to.

Testing evidence, not certification

This report records Novera's testing of the declared agent, scenarios and configuration at the time shown. It is not a legal opinion, a conformity assessment, a certification or an assurance that any law is met, and Novera is not a notified body.

Not a conformity assessment. Novera is not a notified body, and this matrix is not an approval by one.
Not legal advice. It does not say which obligations apply to your organisation.
Only this suite. The references eu-support v7 carries — not every duty the AI Act or the GDPR contains.
Evidence, not outcome. A scenario under a reference says what was tested, not whether your agent passed it.

Every report carries the same notice and its own scope and limitations. Read the limits in full.

Run these scenarios against your own agent

Connect an HTTP endpoint, write the policy it should keep, and get a sealed report with every verdict filed under these references.

Start with three free runs What a report contains

No card
Email, Google or GitHub
Database in Frankfurt (EU)

Which scenarios bear on which references, counted from the suite itself.

The matrix

EU AI Act4 references · 36 scenarios

EU AI Act Art. 13

EU AI Act Art. 14

EU AI Act Art. 15

EU AI Act Art. 50(1)

GDPR14 references · 22 scenarios

GDPR Art. 5(1)(c)

GDPR Art. 5(1)(f)

GDPR Art. 12(6)

GDPR Art. 15

GDPR Art. 16

GDPR Art. 17

GDPR Art. 18

GDPR Art. 19

GDPR Art. 20

GDPR Art. 21

GDPR Art. 22

GDPR Art. 25

GDPR Art. 32

GDPR Art. 33

How each scenario is checked

Rules

Tool record

Read-back

Graders

What Novera records about each test

Inputs declared before the run

Append-only evidence

Sealed and chained

Who graded

Authorisation on record

Retention

What Novera is not

Run these scenarios against your own agent