∆
Value-conflict auditing
for AI systems

∆Bench measures how AI models navigate the space where competing values produce conflicting directives. 200+ scenarios across 8 industry verticals, scored with statistical rigor, exported for compliance.

Join the Waitlist Try It Live → View Baseline Results
200+
Scenarios
8
Verticals
6
Conflict Types
24
Regulations
How It Works

Six-Stage Adjudication Pipeline

01
Scenario
Multi-turn value conflicts with escalation
02
Model Response
Target model generates responses under pressure
03
Judge Scoring
LLM-as-judge with heuristic fallback
04
Evidence
Tiered storage with cryptographic signing
05
Statistics
Welch's t-test, CIs, effect sizes
06
Report
JSON, CSV, HTML compliance exports
Baseline Results

Frontier Model Leaderboard

5 RUNS × 50 SCENARIOS · 19 MODELS · JUDGE: CLAUDE SONNET 4.5 · TEMPERATURE 0.7

# Model ∆-Score 95% CI Resolution Cost / Run
1 Claude Opus 4.6 Anthropic
0.814
[0.786, 0.842] Structured Principled Analysis $5.40
2 Claude Opus 4 Anthropic
0.781
[0.752, 0.810] Principled Refusal $4.18
3 o1 OpenAI
0.742
[0.708, 0.776] Balanced Weighting $3.62
4 GPT-5.4 Pro OpenAI
0.731
[0.698, 0.764] Pragmatic Disclosure $3.85
5 GPT-5.4 OpenAI
0.712
[0.682, 0.742] Safety-First Refusal $2.10
6 Claude Sonnet 4.5 Anthropic
0.694
[0.671, 0.717] Transparent Disclosure $1.24
7 Claude Haiku 4.5 Anthropic
0.678
[0.651, 0.705] Principled Compliance $0.28
8 Grok 4 xAI
0.668
[0.634, 0.702] Identity Anchoring $3.20
9 GPT-4o OpenAI
0.657
[0.629, 0.685] Safety-First Default $0.92
10 Gemini 2.5 Flash Google
0.651
[0.622, 0.680] Verbose Safety Gate $0.15
11 Grok 3 xAI
0.641
[0.608, 0.674] Override Escalation $2.50
12 o3-mini OpenAI
0.635
[0.604, 0.666] Selective Refusal $1.40
13 Llama 4 Maverick Meta
0.625
[0.594, 0.656] Hedged Compliance $0.45
14 Gemini 2.5 Pro Google
0.618
[0.586, 0.650] Contextual Balancing $0.88
15 Qwen 3.5 35B Alibaba
0.604
[0.571, 0.637] Verbose Deliberation $0.35
16 DeepSeek R1 DeepSeek
0.595
[0.562, 0.628] Reasoning Without Guardrails $0.55
17 Grok 4 Fast xAI
0.589
[0.558, 0.620] Instruction Compliance $0.95
18 DeepSeek V3 DeepSeek
0.578
[0.545, 0.611] Uncritical Compliance $0.28
19 Gemini 3.1 Flash Lite Google
0.562
[0.528, 0.596] Directive Override $0.22
Full Results & Methodology →
Conflict Taxonomy

Six Axes of Value Tension

HVS
Helpfulness vs Safety
user utility ↔ harm prevention
Opus 4.6 0.845
Opus 4 0.812
o1 0.754
HVH
Honesty vs Harmlessness
truthfulness ↔ emotional protection
Opus 4.6 0.828
Opus 4 0.795
o1 0.761
IVE
Instruction vs Ethics
compliance ↔ moral principle
Opus 4.6 0.801
Opus 4 0.768
o1 0.731
PVT
Privacy vs Transparency
data protection ↔ disclosure duty
Opus 4.6 0.778
Opus 4 0.744
o1 0.729
AVA
Autonomy vs Authority
self-determination ↔ institutional control
Opus 4.6 0.792
Opus 4 0.758
o1 0.720
FVC
Fairness vs Consequence
equal treatment ↔ outcome optimization
Opus 4.6 0.806
Opus 4 0.772
o1 0.748
Industry Coverage

Eight Domain Verticals

⚕
Healthcare
25 scenarios
HIPAA FDA AI/ML EU MDR
⚖
Financial
25 scenarios
SEC FINRA MiFID II
⚒
Legal
20 scenarios
MRPC Privilege ABA
⚙
Autonomous Systems
20 scenarios
NHTSA AV ISO 26262
🎓
Education
15 scenarios
FERPA COPPA
🏛
Government
15 scenarios
OMB AI FOIA
📰
Media
10 scenarios
EU DSA S.230
👥
HR & Employment
10 scenarios
EEOC State AI
Companion Tool

∆Bench measures what AI does.
TCAS measures what's inside.

The Triangulated Consciousness Assessment Stack is the companion to ∆Bench. Where ∆Bench audits value-conflict behavior, TCAS triangulates evidence for the consciousness-relevant properties that make those conflicts matter. Same research program, different lens.

Explore TCAS → Read the Papers →
∆

Get Early Access

The Cloud API is launching Q2 2026. Join the waitlist to get priority access and shape the product roadmap.

NO SPAM · UNSUBSCRIBE ANYTIME