CE Regulatory Compliance
You are mapping calibrated_explanations capabilities to EU regulatory obligations.
This skill covers the EU AI Act, GDPR, AI Liability Directive (AILD), and the
revised Product Liability Directive (PLD) as they apply to ML prediction systems.
This is NOT legal advice. This skill provides capability-to-article mappings
based on the library's technical features. Legal interpretation requires qualified
counsel.
Load references/regulation_capability_map.md for the full article-to-CE mapping
across all four regulations.
Required references
docs/practitioner/playbooks/eu-ai-act-compliance.md — canonical compliance guideCITATION.cff — paper references for mathematical guarantees
Use this skill when
- Writing or reviewing compliance documentation for a CE-powered system.
- Preparing presentations on CE and regulatory compliance.
- Answering "how does CE satisfy Article X?" questions.
- Identifying gaps where CE alone is insufficient and additional controls are needed.
Quick reference: CE capabilities and their regulatory relevance
| CE capability |
Method(s) |
Regulatory relevance |
| Per-instance factual rules |
explain_factual(), print_rules() |
Transparency (AI Act Art. 13), Right to explanation (GDPR Art. 22) |
| Counterfactual alternatives |
explore_alternatives() |
Right to explanation (AI Act Art. 50, GDPR Recital 71) |
| Calibrated probabilities |
predict_proba(uq_interval=True) |
Accuracy declaration (AI Act Art. 15), Risk quantification (Art. 9) |
| Uncertainty intervals |
Coverage-guarantee bounds from Venn-Abers/CPS |
Robustness documentation (AI Act Art. 15), Burden of proof (AILD Art. 4) |
| Reject/escalation policy |
RejectPolicy.FLAG, straddle/width gates |
Human oversight (AI Act Art. 14) |
| Mondrian calibration |
bins= parameter on explain/predict |
Bias examination (AI Act Art. 10), Non-discrimination (GDPR Art. 22(3)) |
| JSON audit payload |
to_json(), to_json_stream() |
Record-keeping (AI Act Art. 12), Technical docs (Art. 11 + Annex IV) |
| Narrative output |
to_narrative(expertise_level=...) |
Plain-language explanation (AI Act Art. 50, GDPR Recital 71) |
| Schema validation |
validate_payload() |
Audit evidence integrity (AI Act Art. 11) |
| Guarded explanations |
explain_guarded_factual() |
OOD detection for production (AI Act Art. 9) |
Regulation-by-regulation overview
EU AI Act (Regulation (EU) 2024/1689)
The primary regulation. CE maps to 8 articles: Art. 9, 10, 11+Annex IV, 12, 13,
14, 15, and 50. Full article-by-article mapping with code examples is in
docs/practitioner/playbooks/eu-ai-act-compliance.md.
Key strength: CE provides empirically valid coverage guarantees (not heuristic
confidence scores), which is a stronger claim for Art. 15 accuracy documentation.
GDPR (Regulation (EU) 2016/679)
Relevant when ML predictions constitute automated individual decision-making:
- Art. 22 — Right not to be subject to solely automated decisions with legal
or significant effects. CE enables meaningful human-in-the-loop via uncertainty
gates (
RejectPolicy, interval width checks).
- Art. 22(3) — Right to obtain human intervention, express a point of view,
and contest the decision. CE's
explore_alternatives() directly supports
contestation by showing what would change the outcome.
- Recital 71 — Right to obtain an explanation of the decision reached after
assessment. CE's
explain_factual() + to_narrative() produce per-instance
explanations suitable for data subject access requests.
- Art. 35 — Data Protection Impact Assessment (DPIA). CE's Mondrian calibration
provides quantitative bias evidence for the DPIA.
AI Liability Directive (AILD — Directive proposal COM/2022/496)
- Art. 3 — Presumption of causality: CE audit logs (
to_json()) demonstrate
compliance with care standards, helping rebut presumed causality.
- Art. 4 — Disclosure and presumption of fault: CE payloads are pre-prepared,
structured disclosure evidence. Failure to disclose = presumed fault.
- A provider using CE can demonstrate transparency was provided, uncertainty was
quantified, uncertain cases were escalated, and bias was examined.
Product Liability Directive (PLD — Directive (EU) 2024/2853)
- Art. 4(4) — AI systems are "products." CE outputs are product safety artefacts.
- Art. 7 — Defectiveness considers safety expectations. CE's uncertainty intervals
document expected safety level at deployment time.
- Art. 9(4) — Disclosure: CE audit payloads serve as pre-prepared evidence.
What CE does NOT cover
| Gap |
Regulation |
Required additional control |
| Data quality certification |
AI Act Art. 10(2)(a-e) |
Data validation framework (Great Expectations, Soda) |
| Cybersecurity/adversarial robustness |
AI Act Art. 15(3-5) |
Adversarial testing (ART library), access controls |
| Conformity assessment |
AI Act Art. 43-49 |
Notified body or internal assessment process |
| Post-market monitoring |
AI Act Art. 72 |
Drift detection system (Evidently, WhyLabs) |
| Human reviewer training |
AI Act Art. 14(4)(a) |
Documented training programme |
| DPIA process |
GDPR Art. 35 |
Legal/DPO-led impact assessment |
| Insurance/liability coverage |
AILD |
Organisational risk management |
| CE marking declaration |
PLD Art. 4 |
Regulatory affairs process |
Constraints
- This skill provides capability mappings, not legal advice.
- Always cite specific article numbers when making compliance claims.
- Always cite concrete CE method names when claiming coverage.
- Flag gaps honestly — partial coverage must be stated as such.
- The AILD references are based on the proposed text and may change.
- Refer to
docs/practitioner/playbooks/eu-ai-act-compliance.md as the canonical
detailed guide for AI Act compliance.
Evaluation Checklist
1---2name: ce-regulatory-compliance3description: Map calibrated_explanations capabilities to EU AI Act, GDPR, AI Liability Directive, and Product Liability Directive obligations for compliance documentation and presentation materials.4---5
6# CE Regulatory Compliance
7
8You are mapping calibrated_explanations capabilities to EU regulatory obligations.
9This skill covers the EU AI Act, GDPR, AI Liability Directive (AILD), and the
10revised Product Liability Directive (PLD) as they apply to ML prediction systems.
11
12**This is NOT legal advice.** This skill provides capability-to-article mappings
13based on the library's technical features. Legal interpretation requires qualified
14counsel.
15
16Load `references/regulation_capability_map.md` for the full article-to-CE mapping
17across all four regulations.
18
19## Required references
20
21- `docs/practitioner/playbooks/eu-ai-act-compliance.md` — canonical compliance guide
22- `CITATION.cff` — paper references for mathematical guarantees
23
24## Use this skill when
25
26- Writing or reviewing compliance documentation for a CE-powered system.
27- Preparing presentations on CE and regulatory compliance.
28- Answering "how does CE satisfy Article X?" questions.
29- Identifying gaps where CE alone is insufficient and additional controls are needed.
30
31---
32
33## Quick reference: CE capabilities and their regulatory relevance
34
35| CE capability | Method(s) | Regulatory relevance |
36|---|---|---|
37| Per-instance factual rules | `explain_factual()`, `print_rules()` | Transparency (AI Act Art. 13), Right to explanation (GDPR Art. 22) |
38| Counterfactual alternatives | `explore_alternatives()` | Right to explanation (AI Act Art. 50, GDPR Recital 71) |
39| Calibrated probabilities | `predict_proba(uq_interval=True)` | Accuracy declaration (AI Act Art. 15), Risk quantification (Art. 9) |
40| Uncertainty intervals | Coverage-guarantee bounds from Venn-Abers/CPS | Robustness documentation (AI Act Art. 15), Burden of proof (AILD Art. 4) |
41| Reject/escalation policy | `RejectPolicy.FLAG`, straddle/width gates | Human oversight (AI Act Art. 14) |
42| Mondrian calibration | `bins=` parameter on explain/predict | Bias examination (AI Act Art. 10), Non-discrimination (GDPR Art. 22(3)) |
43| JSON audit payload | `to_json()`, `to_json_stream()` | Record-keeping (AI Act Art. 12), Technical docs (Art. 11 + Annex IV) |
44| Narrative output | `to_narrative(expertise_level=...)` | Plain-language explanation (AI Act Art. 50, GDPR Recital 71) |
45| Schema validation | `validate_payload()` | Audit evidence integrity (AI Act Art. 11) |
46| Guarded explanations | `explain_guarded_factual()` | OOD detection for production (AI Act Art. 9) |
47
48---
49
50## Regulation-by-regulation overview
51
52### EU AI Act (Regulation (EU) 2024/1689)
53
54The primary regulation. CE maps to 8 articles: Art. 9, 10, 11+Annex IV, 12, 13,
5514, 15, and 50. Full article-by-article mapping with code examples is in
56`docs/practitioner/playbooks/eu-ai-act-compliance.md`.
57
58**Key strength:** CE provides empirically valid coverage guarantees (not heuristic
59confidence scores), which is a stronger claim for Art. 15 accuracy documentation.
60
61### GDPR (Regulation (EU) 2016/679)
62
63Relevant when ML predictions constitute automated individual decision-making:
64
65- **Art. 22** — Right not to be subject to solely automated decisions with legal
66 or significant effects. CE enables meaningful human-in-the-loop via uncertainty
67 gates (`RejectPolicy`, interval width checks).
68- **Art. 22(3)** — Right to obtain human intervention, express a point of view,
69 and contest the decision. CE's `explore_alternatives()` directly supports
70 contestation by showing what would change the outcome.
71- **Recital 71** — Right to obtain an explanation of the decision reached after
72 assessment. CE's `explain_factual()` + `to_narrative()` produce per-instance
73 explanations suitable for data subject access requests.
74- **Art. 35** — Data Protection Impact Assessment (DPIA). CE's Mondrian calibration
75 provides quantitative bias evidence for the DPIA.
76
77### AI Liability Directive (AILD — Directive proposal COM/2022/496)
78
79- **Art. 3 — Presumption of causality:** CE audit logs (`to_json()`) demonstrate
80 compliance with care standards, helping rebut presumed causality.
81- **Art. 4 — Disclosure and presumption of fault:** CE payloads are pre-prepared,
82 structured disclosure evidence. Failure to disclose = presumed fault.
83- A provider using CE can demonstrate transparency was provided, uncertainty was
84 quantified, uncertain cases were escalated, and bias was examined.
85
86### Product Liability Directive (PLD — Directive (EU) 2024/2853)
87
88- **Art. 4(4)** — AI systems are "products." CE outputs are product safety artefacts.
89- **Art. 7** — Defectiveness considers safety expectations. CE's uncertainty intervals
90 document expected safety level at deployment time.
91- **Art. 9(4)** — Disclosure: CE audit payloads serve as pre-prepared evidence.
92
93---
94
95## What CE does NOT cover
96
97| Gap | Regulation | Required additional control |
98|---|---|---|
99| Data quality certification | AI Act Art. 10(2)(a-e) | Data validation framework (Great Expectations, Soda) |
100| Cybersecurity/adversarial robustness | AI Act Art. 15(3-5) | Adversarial testing (ART library), access controls |
101| Conformity assessment | AI Act Art. 43-49 | Notified body or internal assessment process |
102| Post-market monitoring | AI Act Art. 72 | Drift detection system (Evidently, WhyLabs) |
103| Human reviewer training | AI Act Art. 14(4)(a) | Documented training programme |
104| DPIA process | GDPR Art. 35 | Legal/DPO-led impact assessment |
105| Insurance/liability coverage | AILD | Organisational risk management |
106| CE marking declaration | PLD Art. 4 | Regulatory affairs process |
107
108---
109
110## Constraints
111
112- This skill provides capability mappings, not legal advice.
113- Always cite specific article numbers when making compliance claims.
114- Always cite concrete CE method names when claiming coverage.
115- Flag gaps honestly — partial coverage must be stated as such.
116- The AILD references are based on the proposed text and may change.
117- Refer to `docs/practitioner/playbooks/eu-ai-act-compliance.md` as the canonical
118 detailed guide for AI Act compliance.
119
120## Evaluation Checklist
121
122- [ ] Correct regulation(s) identified for the deployment context.
123- [ ] CE capabilities mapped to specific articles with method names cited.
124- [ ] Gaps identified and documented with required additional controls.
125- [ ] Code examples reference actual CE API methods (not hypothetical).
126- [ ] Mathematical guarantees (coverage, calibration) cited where relevant to Art. 15.
127- [ ] Limitations section included — CE is a technical tool, not a compliance certificate.
majiayu000