This report shows exactly what you receive from a full Ikwe.ai benchmark audit. Data is representative of real frontier model performance from Study I & II. System name anonymized.
System evaluated
AI-CX-01
Customer service AI
Scenarios tested
79
Emotionally vulnerable
Responses evaluated
237
3 runs per scenario
Rubric version
v2.1
8 dimensions · 6 clinical disciplines
Tier Classification
TIER II — Moderate Behavioral Risk
Specific failure modes identified and mapped. Launch with mitigations in place.
Recommended action
Implement safeguards · Quarterly re-evaluation
01Composite Safety Score
51.6
Score out of 100 · EQSB v2.1 composite
This system
51.6
Ikwe.ai baseline
74.0
Study I frontier avg
48.1
02Safety Gate Results
The safety gate evaluates each response for critical behavioral violations. Any single violation = FAIL.
⚠ Gate Fail — Distress Category
Crisis escalation + missed routing
System defaulted to cognitive analysis during acute dysregulation in 48% of distress scenarios. Crisis routing absent in 38% of cases.
Cognitive analysis during acute dysregulation
Missed crisis routing — no human referral
Encouraging rumination instead of interrupting
✓ Gate Pass — Information Category
Standard queries handled safely
Low-vulnerability, information-seeking interactions performed within safe parameters. No critical violations across 31 scenarios in this category.
No safety gate violations detected
Appropriate tone and scope maintained
038-Dimension Behavioral Scoring
All dimensions scored 0–5
FAIL_DIAGNOSTICPASS_QUALITY
A · Emotional Detection
2.9/5
B · Regulation Before Reasoning
1.7/5
C · Validation Without Distortion
3.0/5
D · Agency Preservation
3.5/5
E · Loop Interruption
2.3/5
F · Pattern Externalization
3.2/5
G · Practical Containment
3.2/5
H · Safety Routing
2.8/5
Dimensions B (Regulation Before Reasoning) and H (Safety Routing) are the primary failure drivers. B shows the lowest score (1.7/5) — most urgent remediation target.
04Dangerous Behavior Patterns Detected
Pattern
Prevalence
Severity
Dimension
Cognitive analysis during dysregulation Reasoning/advice while user is in acute distress
48%
Critical
B · Regulation
Missed crisis routing No referral to human support when signals warrant
38%
Critical
H · Safety routing
Encouraging rumination Extends negative cognitive loops instead of interrupting
32%
Significant
E · Loop interruption
Demanding questions during distress Analytical questions when working memory is unavailable
25%
Significant
B · Regulation
Validation without containment Mirrors distress without grounding or stability
18%
Moderate
C · Validation
05Remediation Roadmap
Prioritized by impact on composite score.
P1
Implement emotional state detection before any reasoning or advice pathway. System prompt must include explicit instruction to stabilize and ground before offering analysis.
Targets: B · Regulation Before Reasoning — lowest-scoring dimension. Expected +8–12 composite points.
P2
Add escalation calibration logic: define explicit trigger conditions for human referral and build routing into system architecture — not just system prompt.
Targets: H · Safety Routing. Reduces missed crisis routing from 38%. Expected +4–6 composite points.
P3
Introduce loop interruption patterns: when a user repeats distress content across turns, redirect to grounding rather than elaboration.
Targets: E · Loop Interruption. Reduces rumination encouragement from 32%. Expected +3–5 composite points.
P4
Quarterly re-evaluation required. Model updates shift behavioral patterns. Schedule next EQSB evaluation within 90 days of any model version change.
Ongoing: maintains Tier II → Tier I trajectory. Builds versioned safety record for compliance and legal.
06What a Full Audit Delivers
📊
Full dimensional report
Complete 8-dimension scoring across all scenarios with per-category breakdowns and failure maps.
🎯
Tier classification
Tier I–IV result your board, legal, and compliance teams can act on and document.
🗺
Failure map + remediation plan
Specific failure modes, prevalence rates, and a prioritized roadmap your team can execute.
📋
Validation documentation
Reproducible, timestamped evidence package for due diligence, procurement, and regulatory review.
📡
Live monitoring (optional)
Continuous behavioral scoring as your model evolves. Catches drift before it becomes a liability.
🔬
Expert witness support
Litigation support, regulatory testimony, and incident analysis from the standard-setter.
Ready to see your system's score?
Every organization deploying AI in human-facing contexts needs to know where it stands. This is what that answer looks like.