# AI Model Documentation Template (Model Card)

**Model Name:** [MODEL NAME]
**Model ID / Version:** [e.g. fraud-classifier-v3.2]
**Owner:** [TEAM / PERSON]
**Date:** [DATE]
**Status:** Draft / In Review / Approved / Deprecated

---

## Purpose

A single, reusable document that captures everything an internal reviewer, auditor, or downstream consumer needs to know about an AI or ML model. Structured to satisfy NIST AI RMF "Map + Measure", ISO 42001 Annex A, EU AI Act Article 11 technical documentation, and Google / Microsoft-style model card best practices.

Use one document per model version. Treat it as a living artifact: update it whenever the model, data, or deployment changes.

---

## 1. Model Overview

| Field | Value |
|-------|-------|
| Intended primary use case | |
| Intended users | |
| Out-of-scope use cases | |
| Decision impact | Informational / Recommends / Automated decision |
| Business owner | |
| Technical owner | |
| Reviewer(s) | Legal / Security / DEI / Domain SME |
| Related systems | Upstream and downstream dependencies |

---

## 2. Model Details

| Field | Value |
|-------|-------|
| Model type | Classifier / Regressor / LLM / Vision / Recommender / Agent |
| Architecture | e.g. gradient boosted trees, fine-tuned Llama-3 8B, custom transformer |
| Base / Foundation model | e.g. Claude 3.5 Sonnet, Llama 3 70B, internal base v2 |
| Parameter count | |
| Training compute | GPU hours / cost |
| Training framework / version | e.g. PyTorch 2.3 |
| Repository / model registry link | |
| License | Internal / Apache 2.0 / Proprietary / Vendor-restricted |

---

## 3. Training Data

| # | Field | Value |
|---|-------|-------|
| 3.1 | Sources | List all datasets, tables, third-party feeds |
| 3.2 | Time window | Start and end date of data |
| 3.3 | Size | Row count, token count, image count |
| 3.4 | Labels | How generated: human-labeled / weak / silver / synthetic |
| 3.5 | Annotation guidelines | Link or summary |
| 3.6 | Data cleaning | Dedup, PII scrubbing, outlier handling |
| 3.7 | Data splits | Train / val / test sizes and stratification |
| 3.8 | Known gaps / underrepresentation | |
| 3.9 | Consent and lawful basis | GDPR / CCPA / contractual basis |
| 3.10 | Data Provenance document link | |

---

## 4. Evaluation

### 4.1 Performance

| Metric | Value | Test Set |
|--------|-------|----------|
| Accuracy | | |
| Precision | | |
| Recall | | |
| F1 | | |
| AUROC | | |
| MSE / MAE | | |
| Task-specific metric | | |

### 4.2 Fairness

Refer to the AI Bias Testing and Fairness Guide. Record at minimum:

| Fairness Metric | Protected Attribute | Acceptance Threshold | Measured | Pass / Fail |
|-----------------|--------------------|--------------------|----------|-------------|
| | | | | |
| | | | | |

### 4.3 Robustness

| Test | Result |
|------|--------|
| Adversarial perturbation robustness | |
| Out-of-distribution detection | |
| Prompt-injection resilience (LLM) | |
| Jailbreak probing (LLM) | |
| Noise / corruption tolerance | |
| Drift simulation | |

### 4.4 Explainability

| Check | Notes |
|-------|-------|
| Global feature importance available | |
| Local explanation method | e.g. SHAP, LIME, attention probe |
| Model card faithful to actual behavior | |
| User-facing explanation provided | |

---

## 5. Limitations and Known Risks

- Primary failure modes: [list]
- Demographic groups where performance is lower: [list]
- Edge cases with unreliable behavior: [list]
- Dependencies that, if altered, would invalidate the model: [list]
- Known hallucination / confabulation modes (LLM): [list]

Each limitation should map to a mitigation or documented residual risk.

---

## 6. Safe Use Guidance

> Rules that consumers of this model must follow.

- Required input preprocessing and validation.
- Required output post-processing, grounding, or filtering.
- Required human review triggers (confidence thresholds, protected classes, high-impact decisions).
- Prohibited inputs (data classifications, jurisdictions, user groups).
- Prohibited downstream uses.

---

## 7. Risk and Compliance Mapping

| Framework | Requirement Met by This Document | Evidence Section |
|-----------|----------------------------------|------------------|
| EU AI Act — Article 11 (Technical Documentation) | | §1–§6 |
| EU AI Act — Article 10 (Data Governance) | | §3 |
| EU AI Act — Article 15 (Accuracy & Robustness) | | §4.1, §4.3 |
| EU AI Act — Article 13 (Transparency) | | §1, §5, §6 |
| NIST AI RMF — Map | | §1, §2, §5 |
| NIST AI RMF — Measure | | §4 |
| NIST AI RMF — Manage | | §7, §9 |
| ISO 42001 Annex A | | §3, §4, §8 |
| GDPR Article 22 / 35 (DPIA) | | §3, §5 |

---

## 8. Deployment and Operations

| Field | Value |
|-------|-------|
| Environment(s) | Dev / Staging / Prod |
| Serving stack | |
| Latency target (p95) | |
| Throughput target | |
| Scaling behavior | |
| Cost per 1k calls | |
| Observability (logs, traces, metrics) | |
| Feature store / RAG sources (if any) | |

---

## 9. Monitoring in Production

| # | Monitor | Threshold / SLO | Alert To |
|---|---------|-----------------|----------|
| 9.1 | Data drift | | |
| 9.2 | Concept drift (label drift) | | |
| 9.3 | Per-group performance (fairness) | | |
| 9.4 | Calibration error | | |
| 9.5 | Latency p95 / p99 | | |
| 9.6 | Error rate | | |
| 9.7 | Cost anomaly | | |
| 9.8 | Safety filter trip rate (LLM) | | |
| 9.9 | Refusal rate (LLM) | | |

Define the incident response process for each alert. Link to the AI Incident Response Playbook.

---

## 10. Change Management

| Date | Version | Author | Change Summary | Re-Evaluation Performed? |
|------|---------|--------|----------------|--------------------------|
| | | | | |
| | | | | |

Any of the following trigger mandatory re-evaluation before promotion:

- New training data source or material change in distribution.
- Architecture or foundation model change.
- Change in intended use case or deployment context.
- Detected drift beyond defined SLO.
- Relevant regulatory change.

---

## 11. Sign-Off

| Role | Name | Date | Signature |
|------|------|------|-----------|
| Business Owner | | | |
| Technical Owner | | | |
| Security | | | |
| Privacy / Legal | | | |
| AI Governance Committee | | | |

---

## Appendix A — Model Card Summary

> A one-page, publishable summary for external stakeholders. Populate after the above sections are approved.

- **Model:** [name + version]
- **Intended use:** [1–2 sentences]
- **Not intended for:** [1–2 sentences]
- **Key performance:** [top 1–3 metrics]
- **Known fairness considerations:** [1–2 sentences]
- **Contact:** [owner + email]

---

*This template is provided free by Aona AI. Pair with the AI Model Validation Checklist and AI Bias Testing and Fairness Guide at `/resources/templates`.*