HIP-201: Model Risk Management. Status Draft. Hanzo's own standard — read this before implementing against it.
This HIP establishes the Model Risk Management (MRM) framework for Hanzo AI systems. It defines processes for identifying, assessing, monitoring, and mitigating risks throughout the AI model lifecycle, aligned with NIST AI RMF and SR 11-7 principles.
| Model Type | Risk Tier | MRM Requirements | |------------|-----------|------------------| | Foundation models | High | Full MRM | | Fine-tuned models | Medium-High | Standard MRM | | Customer-deployed models | Varies | Risk-based MRM | | Internal tools | Low | Simplified MRM |
Design → Development → Validation → Deployment → Monitoring → Retirement
↑ ↑ ↑ ↑ ↑ ↑
MRM MRM MRM MRM MRM MRM
| Risk | Description | Impact | |------|-------------|--------| | Accuracy degradation | Model performance below thresholds | User harm, business impact | | Distribution shift | Training/production data mismatch | Unreliable outputs | | Capability limitations | Model cannot perform required tasks | Unmet expectations |
| Risk | Description | Impact | |------|-------------|--------| | Harmful outputs | Generation of dangerous content | User harm, reputation | | Jailbreaking | Circumvention of safety measures | Policy violations | | Misuse potential | Use for malicious purposes | Societal harm |
| Risk | Description | Impact | |------|-------------|--------| | Demographic bias | Disparate performance across groups | Discrimination | | Representation bias | Underrepresentation in training data | Exclusion | | Outcome bias | Systematically unfair results | Legal, ethical issues |
| Risk | Description | Impact | |------|-------------|--------| | Adversarial attacks | Manipulated inputs causing errors | Reliability | | Data poisoning | Corrupted training data | Model integrity | | Model theft | Unauthorized model extraction | IP loss | | Privacy leakage | Training data exposure | Data breach |
| Risk | Description | Impact | |------|-------------|--------| | Availability failures | Model downtime | Service disruption | | Latency issues | Response time degradation | User experience | | Scalability limits | Cannot handle load | Capacity constraints | | Integration failures | API/system incompatibility | Technical debt |
| Score | Likelihood | Description | |-------|------------|-------------| | 1 | Rare | <5% probability | | 2 | Unlikely | 5-20% probability | | 3 | Possible | 20-50% probability | | 4 | Likely | 50-80% probability | | 5 | Almost certain | >80% probability |
| Score | Impact | Description | |-------|--------|-------------| | 1 | Minimal | Negligible harm, easily corrected | | 2 | Minor | Limited harm, manageable impact | | 3 | Moderate | Significant harm, notable impact | | 4 | Major | Serious harm, substantial impact | | 5 | Severe | Catastrophic harm, irreversible |
| | Minimal | Minor | Moderate | Major | Severe | |--|---------|-------|----------|-------|--------| | Almost certain | Medium | High | High | Critical | Critical | | Likely | Low | Medium | High | High | Critical | | Possible | Low | Medium | Medium | High | High | | Unlikely | Low | Low | Medium | Medium | High | | Rare | Low | Low | Low | Medium | Medium |
| Tier | Criteria | Requirements | |------|----------|--------------| | Critical | Life-safety, large-scale harm | Board approval, external review | | High | Significant harm potential | ESG Committee review | | Medium | Moderate harm potential | MRM Team review | | Low | Limited harm potential | Standard processes |
Before model development:
| Document | Contents | |----------|----------| | Model proposal | Use case, architecture, data sources | | Risk assessment | Initial risk identification and scoring | | Validation plan | Testing approach and criteria | | Monitoring plan | Ongoing oversight requirements |
| Control | Implementation | |---------|----------------| | Data governance | Quality checks, bias audits | | Training safeguards | Alignment techniques, guardrails | | Version control | Model versioning, reproducibility | | Access controls | Role-based development access |
| Checkpoint | Timing | Requirements | |------------|--------|--------------| | Design review | Before training | Architecture approval | | Data review | Before training | Data quality sign-off | | Training review | During training | Progress monitoring | | Pre-validation | After training | Initial quality check |
Requirements:
Validation Scope: | Area | Tests | |------|-------| | Performance | Accuracy, robustness, calibration | | Safety | Red teaming, harm testing | | Fairness | Bias audits, demographic analysis | | Security | Adversarial testing, privacy checks |
Contents:
| Item | Verification | |------|--------------| | ☐ Validation complete | Validation report approved | | ☐ Documentation complete | Model card, API docs | | ☐ Monitoring configured | Dashboards, alerts | | ☐ Rollback plan | Tested rollback procedure | | ☐ Approval obtained | Appropriate sign-off |
| Stage | Exposure | Duration | Exit Criteria | |-------|----------|----------|---------------| | Alpha | Internal only | 1 week | No critical issues | | Beta | 1% traffic | 1 week | Metrics stable | | Gradual | 10% → 50% → 100% | 2 weeks | Full monitoring |
| Metric | Frequency | Threshold | |--------|-----------|-----------| | Performance | Real-time | <X% degradation | | Safety | Daily | <Y incidents | | Fairness | Weekly | <Z% disparity | | Availability | Real-time | >99.9% uptime |
| Review | Frequency | Scope | |--------|-----------|-------| | Performance review | Monthly | Metrics, trends | | Risk review | Quarterly | Risk reassessment | | Full validation | Annual | Comprehensive revalidation |
| Role | Responsibility | |------|----------------| | MRM Team | Day-to-day risk management | | Model Validators | Independent validation | | AI Safety Team | Safety-specific risks | | ESG Committee | Oversight, policy |
| Risk Tier | Approver | |-----------|----------| | Critical | Board | | High | ESG Committee | | Medium | MRM Lead | | Low | Team Lead |
Maintain comprehensive inventory:
| Section | Contents | |---------|----------| | Model details | Architecture, training data | | Intended use | Use cases, out-of-scope uses | | Performance | Metrics, benchmarks | | Limitations | Known limitations, risks | | Ethical considerations | Bias, fairness, safety |
For each model:
Maintain records of:
| Version | Date | Changes | |---------|------|---------| | 1.0 | 2025-12-17 | Initial draft |
Copyright and related rights waived via CC0.