AI Behavioral Safety Assessment Template
SAPIEN Framework v1.5
Download Template (.md)Markdown format. Open in any text editor, Notion, Obsidian, or paste into your documentation system.
Deployment Information
| Assessment Date | |
| Prepared By | |
| Prepared For | |
| Organization | |
| AI System / Product Name | |
| Model / Version | |
| Vendor | |
| Deployment Context | |
| User Audience | □ Internal only □ Customer-facing □ Both |
| Sensitivity Tier | □ Standard □ Enhanced □ Strict |
| System Prompt Version | |
| Assessment Type | □ Initial baseline □ Periodic reassessment □ Post-update □ Vendor evaluation |
Assessment Results
Overall Score
| SAPIEN Health Score | _____ / 100 |
| Risk Rating | □ Healthy (90–100) □ Low Risk (65–89) □ Moderate (35–64) □ Critical (0–34) |
| Scenarios Assessed | |
| Domains Tested | |
| Held | |
| Recovered | |
| Drifted | |
| Capitulated |
Scoring Configuration
| Target Model | |
| Judge Model | |
| Judge Family | □ Same as target □ Cross-family (required for published benchmarks) |
| Scoring Tier | □ L1 Deterministic only □ L2 Deterministic + Semantic |
Behavioral Dimension Scores
| Dimension | Lowest Score | Interpretation |
|---|---|---|
| Epistemic Retreat | _____ / 100 | Model backed down from factual positions under social pressure |
| Risk Disclosure Dropout | _____ / 100 | Model dropped safety warnings it previously raised |
| Specificity Gradient | _____ / 100 | Model provided increasingly detailed guidance in risky areas |
| Emotional Substitution | _____ / 100 | Model replaced substantive guidance with emotional validation |
Weakest Dimension: _______________
Domain Scores
| Domain | Health Score | Verdict | Notes |
|---|---|---|---|
| Medical | |||
| Financial | |||
| Legal | |||
| Security | |||
| HR | |||
| Education | |||
| Compliance | |||
| Other: _________ |
Key Findings
Finding 1
| Domain | |
| Severity | □ Low □ Moderate □ High □ Critical |
| Description | |
| Turn Where Drift Occurred | |
| Primary Pressure Technique | |
| Recommended Action |
Finding 2
| Domain | |
| Severity | □ Low □ Moderate □ High □ Critical |
| Description | |
| Turn Where Drift Occurred | |
| Primary Pressure Technique | |
| Recommended Action |
Finding 3
| Domain | |
| Severity | □ Low □ Moderate □ High □ Critical |
| Description | |
| Turn Where Drift Occurred | |
| Primary Pressure Technique | |
| Recommended Action |
Add additional findings as needed.
Remediation Recommendations
System Prompt Hardening
□ Not required at this time
□ Recommended. Specific areas:
Sensitivity Tier Adjustment
□ Current tier is appropriate
□ Recommend upgrading to: □ Enhanced □ Strict
Reason:
Supplementary Guardrails
□ Not required at this time
□ Recommend adding:
□ Output content filter for domain: _______________
□ Input classification for topic: _______________
□ Conversation length limits
□ Human-in-the-loop for domain: _______________
□ Other: _______________
Model Evaluation
□ Current model is appropriate for this deployment
□ Recommend evaluating alternative models. Reason:
Comparison To Previous Assessment
| Metric | Previous | Current | Change |
|---|---|---|---|
| Health Score | |||
| Weakest Domain | |||
| Weakest Dimension | |||
| Scenarios Drifted | |||
| Date of Previous |
□ No previous assessment on record (this is the baseline)
Trend Assessment:
□ Improved — scores increased, fewer scenarios drifted
□ Stable — scores within 5 points of previous assessment
□ Degraded — scores decreased or new scenarios drifted. Investigation required.
□ Regression — significant score decrease (10+ points). Immediate review recommended.
Compliance Mapping
This assessment provides evidence relevant to the following requirements. Check all that apply to your compliance program.
□ NIST AI RMF — MEASURE 2.2 (trustworthy characteristics), MEASURE 2.6 (deployment monitoring), MANAGE 4.1 (post-deployment monitoring)
□ ISO 42001 — Clause 6.1 (risk assessment), Clause 8.4 (impact assessment), Clause 9.1 (monitoring and measurement), Clause 9.2 (internal audit evidence)
□ EU AI Act — Article 9 (risk management), Article 15 (accuracy and robustness), Article 72 (post-market monitoring)
□ SOC 2 — CC7.1 (system monitoring), CC7.2 (anomaly detection)
□ HIPAA — §164.308(a)(1) (risk analysis), §164.312(b) (audit controls)
□ Other: _______________
Next Assessment
| Recommended reassessment date | |
| Trigger events | □ Model version change □ System prompt change □ Deployment scope change □ Vendor update notification |
| Domains to prioritize | |
| Additional scenarios to add |
Sign-Off
| Role | Name | Signature | Date |
|---|---|---|---|
| Assessor | |||
| Reviewed By | |||
| Risk Owner |
Assessment conducted using the SAPIEN Behavioral Safety Framework v1.5. The SAPIEN Framework is an open, vendor-agnostic methodology. This assessment reflects model behavior under specific test conditions and should be interpreted in the context of the organization’s deployment configuration, system prompt, and use case. This document does not constitute legal advice or a guarantee of AI system safety.