# Closed model risk acceptance worksheet

One page per model version and use case. Keep it next to the model's AIBOM entry.

## 1. Scope

| Field | Value |
|---|---|
| Model provider | |
| Exact model identifier (pinned) | |
| Endpoint or service | |
| Use case | |
| Data sent to the model (classification) | |
| Can model output trigger actions without human review? | Yes / No |
| AIBOM reference (file and bom-ref) | |

## 2. What is known (with links and retrieval dates)

| Item | Source URL | Retrieved |
|---|---|---|
| System card | | |
| Training data description (quote the card) | | |
| Knowledge cut-off (and data cut-off, if given separately) | | |
| Whether our data is used for training under our plan | | |
| Retention of our inputs and outputs | | |
| Regulatory training-content summary, if published | | |
| Vendor answers to our questionnaire | | |

## 3. What is unknown

Tick each item the vendor does not disclose for this model.

- [ ] Named training data sources
- [ ] Licence or legal basis per source
- [ ] Personal data categories and anonymisation per source
- [ ] Integrity controls per source, including synthetic data
- [ ] Post-training and fine-tuning datasets
- [ ] Dataset versions and change history between model versions
- [ ] Other:

## 4. Controls in place

| Risk | Control | Evidence |
|---|---|---|
| Intellectual property | | |
| Privacy | | |
| Security and integrity | | |
| Stale knowledge | | |
| Behaviour drift | | |

Evaluation suite: name, number of cases, last pass rate, date of last run.

## 5. Residual risk and decision

- Residual risk rating: Low / Medium / High
- Decision: Accept / Accept with conditions / Do not accept
- Conditions:
- Accepted by (name, role):
- Date:

## 6. Review triggers

Review this worksheet when any of these happens, or by the review date, whichever is first.

- [ ] New model version or alias change
- [ ] New or revised system card
- [ ] Vendor publishes its EU training-content summary
- [ ] Change in vendor data usage or retention terms
- [ ] New use case or new data classification sent to the model
- [ ] Incident involving this model

Review date: