Governance documentation

Governance Documentation

You cannot deploy an AI system without governance documentation. This skill generates the minimum required artefacts. Do not ship without completing every item.

What Must Exist Before Deployment

Document	Required For	Generated By
Model Card	All AI systems	This skill — Step 1
Risk Assessment	Tier 2+ systems	This skill — Step 2
ATRS Entry	UK public sector	This skill — Step 3
Incident Response Plan	All production systems	This skill — Step 4
Monitoring Plan	All production systems	This skill — Step 5

---

Step 1: Generate the Model Card

Run the model card generator:

python skills/governance-documentation/scripts/generate_model_card.py \
  --name "Your System Name" \
  --output docs/model-card.md

Or fill in this template manually:

# Model Card: [System Name]

## System Details
- **Name:** 
- **Version:** 
- **Date:** 
- **Type:** [Classification / Regression / NLP / Generative / Agent]
- **Owner:** 
- **Contact:** 

## Intended Use
- **Primary use:** 
- **Intended users:** 
- **Out-of-scope uses:**

## Training Data
- **Source:** 
- **Size:** 
- **Date range:** 
- **Known limitations:** 

## Performance
| Metric | Overall | Group A | Group B | Group C |
|---|---|---|---|---|
| Accuracy | | | | |
| F1 | | | | |
| FPR | | | | |

## Limitations
- 
- 

## Ethical Considerations
- Potential for harm: 
- Mitigations: 

## Human Oversight
- Who reviews outputs: 
- Override mechanism: 
- Override rate (target): 

---

Step 2: Complete the Risk Assessment

Answer every question. No blanks.

## AI Risk Assessment — [System Name]

### Impact Classification
- Affected population: 
- Decision reversibility: [Easily reversed / Difficult / Irreversible]
- Vulnerable groups affected: [Yes/No — specify]
- Maximum consequence of error: 
- Risk Tier: [1 / 2 / 3 / 4]

### Data Risks
- Personal data processed: [Yes/No]
- Special category data: [Yes/No — specify]
- DPIA completed: [Yes/No — date]
- Data minimisation applied: [Yes/No]

### Model Risks
- Bias assessment completed: [Yes/No — date]
- Fairness tests passing: [Yes/No]
- Known failure modes:
  1. 
  2. 
- Mitigation for each:
  1. 
  2. 

### Operational Risks
- What happens if system is unavailable: 
- Manual fallback process: 
- Time to restore service: 
- Monitoring in place: [Yes/No — describe]

### Residual Risk
- Overall residual risk: [Low / Medium / High / Critical]
- Accepted by (SRO name and date): 

---

Step 3: ATRS Entry (UK Public Sector Only)

If deploying in UK public sector context, complete this before go-live:

## ATRS Entry — [System Name]

**Tool name:** 
**Owning organisation:** 
**Description:** [Plain English — what does this tool do, and in which decisions?]
**Phase:** [In development / Deployed / Decommissioned]
**Scope:** [Which decisions or processes does it apply to?]
**Type of AI:** 
**Supplier:** 
**Date first deployed:** 
**Review date:** 

**Human oversight:**
[Describe exactly what human review happens before any decision is acted on]

**Equality impact:**
[Was an EIA completed? What did it find? What was done about it?]

**Data used:**
[What data does the system use? Is any of it personal or sensitive?]

**Where to find more information:**
[Link to model card, contact details]

Publish this at: [your-org-website]/algorithmic-transparency

---

Step 4: Incident Response Plan

Complete this template. Every field is mandatory:

## AI Incident Response Plan — [System Name]

### Severity Levels
| Level | Description | Response Time | Escalation |
|---|---|---|---|
| P1 — Critical | System causing active harm to citizens | 1 hour | SRO + Director immediately |
| P2 — High | Significant performance degradation | 4 hours | SRO within 2 hours |
| P3 — Medium | Fairness threshold exceeded | 24 hours | Team lead + SRO |
| P4 — Low | Anomalous output detected | 72 hours | Team lead |

### First Response Checklist
- [ ] Assess severity level
- [ ] Log incident in incident register with timestamp
- [ ] Notify appropriate contacts (see escalation matrix below)
- [ ] Decide: continue / pause / shut down system
- [ ] If pausing: activate manual fallback process
- [ ] Preserve all logs — do not delete or overwrite

### Emergency Shutdown Procedure
1. [Exact steps to stop the system — must be executable in <15 minutes]
2. 
3. 

### Escalation Matrix
| Severity | Primary Contact | Secondary | Regulator Notification? |
|---|---|---|---|
| P1 | [Name, phone] | [Name, phone] | ICO within 72h if personal data breach |
| P2 | [Name, phone] | [Name, phone] | Assess case by case |
| P3 | [Name, phone] | | No |

### Post-Incident Review
- All P1/P2 incidents: mandatory post-incident review within 5 business days
- Output: root cause analysis + remediation plan + updated risk assessment
- Share with governance board within 10 business days

---

Step 5: Monitoring Plan

## Monitoring Plan — [System Name]

### Metrics to Monitor
| Metric | Baseline | Warning Threshold | Critical Threshold | Frequency |
|---|---|---|---|---|
| Prediction accuracy | | | | Daily |
| Demographic parity gap | | >0.05 | >0.10 | Weekly |
| Override rate | | <5% or >50% | | Weekly |
| System availability | 99.5% | <99% | <95% | Real-time |
| Error rate | | | | Daily |

### Monitoring Owner
- Primary: [Name, role]
- Secondary: [Name, role]

### Review Cadence
- Automated alerts: real-time
- Weekly dashboard review: [day and time]
- Monthly governance report: [date]
- Annual full review: [month]

### Model Retraining Triggers
- [ ] Accuracy drops below [threshold]
- [ ] Fairness threshold exceeded for [N] consecutive weeks
- [ ] Significant change in input data distribution
- [ ] Policy or legislation change affecting the use case
- [ ] Annual scheduled review

---

Completion Checklist

• Model card written and stored in docs/model-card.md
• Risk assessment complete and signed off by SRO
• ATRS entry drafted (publish before or on go-live date)
• Incident response plan complete — emergency shutdown tested
• Monitoring plan in place with named owner
• All documents committed to version control
• Documents location communicated to governance board

You may not deploy to production until all items are checked.

Now run alignment-review as the final check before go-live.